> Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected (I guess ultimately they're all the same thing but still) OR a few may have had a smidge of Alpha but as soon as you add transaction costs it all disappears.
> I should caveat that I was a profitable trader at multiple Tier-1 US banks so I can say with confidence that I made a decent attempt of building whatever the author was trying to get at.
> Almost every instrument is mean-reverting on short timelines and trending on longer timelines.
i.e. he confirms the momentum factor, which isn't surprising since there's more solid evidence for it than anything else, going back hundreds of years.
He doesn't say what fundamental factors he looked at, so it's possible that value, size, and profitability/quality would hold up as well. All those have been studied pretty extensively in academia, in papers going back decades. The author took only a fairly random sampling of recent papers.
All the finance folks I know use "fundamentals" to mean some class of attributes, that they all know what they are. Like when I say "GPL3" to someone who has a-priori knowledge.
Not exactly sure but I think the "fundamentals" are top-line revenue, unit cost, unit margin, yoy-growth, EBTDA, cash-on-hand, default-alive - and likely the ratios those values produce.
These sort of papers don't look at all the fundamentals at once. They'll isolate one factor, or a small related group, and see if it's predictive.
Two simple factors that are well supported: on a risk-adjusted basis, stocks of small companies do better than big ones, and value stocks do better than growth stocks. ("Value" means the stock is cheap relative to some simple fundamental measure like the company's book value.)
A simpler way to think of it is any data that comes off the Income Statement, Balance Sheet, or Cash Flow Statement - or can be directly calculated using such data.
Please don't downvote this comment. There's a huge difference between short term trading and investing. And yes for short term trading (minutes/hours/few days) fundamentals matter rather litte - after all you can be a successful trader of Bitcoin, which has pretty much no fundamentals. Except on volatility events like macro data releases, earnings, FED decisions, surprise news (looking at you GE) etc.
Having a small impact is not the same a having zero impact.
Traders operate over longer timeframe even if they are holding an individual stock for minutes there is a limited pool of stocks. Keep playing the game and longer term impacts add up.
But do the fundamentals of an individual investment matter in the short term compared to the fundamentals of the market as a whole? I would suspect that the “longer term impacts add[ing] up” would just be market health. Day-trading randomly-picked stocks with random buys and sells is a Markov approximation of an index fund :)
I think that depends on how stocks are chosen when you are day trading. Limiting things to high volatility stocks for example creates a bias in your index fund approximation.
I know as little about this as you do; my own gathering is that “fundamentals” are any signal you can get from a public company’s mandated quarterly reporting. So, every piece of data on the published balance sheet, plus maybe any reported board actions.
Interesting but it would be nice if the author would, you know, write up his/her own detailed analysis with replication steps and post on arxiv or something.
I got the feeling this was more of a case of "I did all this work for myself. Nothing useful came up, so here's what I found." The author may be ok with spending an hour sharing the findings, but doesn't want to spend more time than that.
Maybe he could work with someone else to get it to a publication stage? It's super interesting, and combats a well-known problem that you get a lot more out of publishing positive cases (we found a correlation) vs negative cases (we found no correlation), although they're both valuable knowledge.
Look, I know writing stuff up sucks. But it could be a great opportunity to learn a lot of things from a very knowledgable person. With the right prof, it could be a great undergrad project.
Which honestly feels like a reasonable position. Though the untrusting, conspiracy-minded part of my brain wonders if I did review 130 different research papers and found 1 that worked if I'd keep it a secret and just tell the world that they were all crap.
Except plenty useful came up. "Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected" out of 130+ is a significant result. The media would jump on this, and they just might even without any proof of work.
I was referring to the word "useful" within the context of the author's hypothetical goal, not the parent poster's (i.e. a strategy useful for making money opposed to sharing knowledge).
This is one of those cases where I would have guessed this would be the case, but it's nice that somebody else spent their time to verify, since I'm unwilling to spend my time to do so. Also nice that they shared their experience with the rest of us.
If it worked, it wouldn't be published, or at least not until it stopped working.
> Also nice that they shared their experience with the rest of us.
Except they didn't share the results with the rest of us. I know you said "experience" not "results", but when disproving papers, the least you can do is write down three sentences about each paper as you go along reproducing them, noting what you are seeing, perhaps with a snapshot (just a zip file or so) of the code. This is calling a whole field nonsense (that everyone expects to be full of nonsense) without giving enough evidence for anyone else to dispute your claims.
This is unsurprising. P likely is not equivalent to NP [0], and predicting the market is NP-hard [1]. It's nice to see empirical work in the field, though, and especially nice to see reproductions of published papers.
Edit for downvoters and repliers: If enough market participants are irrational, then it can still be possible for people to predict other people, instead of the market, and make money that way.
NP-hardness indeed doesn't rule out heuristic approaches, but experience with 3-SAT and other NP-complete problems suggest that there will be arbitrarily bad times, and that in those times, the amount of loss can be exponential in the length of time that the heuristic poorly predicts the market.
Philip Maymin seems like a serious guy... but that EMH ↔ P=NP paper is absolutely not even remotely a proof. Was genuinely very curious and it's at best an intuition. Some claims, e.g. Knapsack and 3SAT are (almost?) isomorphic to the efficient market hypothesis, are pretty bold. And the justification is hand-wavy at best.
It's not just hand-wavy. It's outright wrong. The size of the search space isn't just exponential in the amount of history -- that's the number of possible histories. The number of mappings from histories to positions is doubly exponential, so they can't even be uniquely described in less than exponential space.
You are talking about writing an algorithm that has a 100% accuracy on an NP-hard problem, and taking the impossibility of this to discard an approach that may yield 58% accuracy.
Warren Buffet is an interesting case, because his approach seems largely to consist of ways to reduce the N of what he is considering: choose a few bets, bet big, hold for the long term, choose bets where you already understand the market. Sampling, rarely recalculating and selecting subsets of N where the related data is already in cache are all well-understood techniques for addressing computationally-hard problems.
Your second link doesn't say predicting the market is NP-hard. It says the opposite: that the market is only efficient if P=NP.
According to the paper, if you don't believe that P=NP then you believe that the market is inefficient, which means there's profit to be made. The paper even suggests how.
Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected (I guess ultimately they're all the same thing but still) OR a few may have had a smidge of Alpha but as soon as you add transaction costs it all disappears.
I could have told you that without testing. If anyone had a lucrative strategy would they disclose it in a paper to the general public? I think not.
CAPM, Fama French Three Factor, or five factor? Are you serious? They don’t work anymore but they did at one point. The foundation of modern finance is built on published Chicago school papers.
From what I've read, they don't work quite as well but they do still seem to work. Economists have spent a lot of time trying to figure out why. Generally, they come up with behavioral explanations, like recency bias, and structural ones, like agency issues.
Why is this getting downvoted? The original post is clearly marketing clickbait. It's trivially true that most "predict the stock market" papers are going to be bunk.
I think people aren't skeptical about this because it lines up with what they already expect. That's the case for me.
If the guy is lying and it's all made up, I'd still wager that if you assembled a collection of 130 papers on this topic, at most 10% would be valid [1]. So even if he's wrong, he's probably not wrong.
[1] I have an academic background, and I think invalid papers make it through peer review constantly. Peer review is not what people think it is.
Wouldn't any given approach rapidly lose efficacy as soon as its published?
I would even guess that a paper being published means that, at the point the paper started to be written, its alpha had already decreased to zero. Otherwise the writers of the paper would still be using that approach. That's how it can appear to provide no value even if you extrapolate it back in time.
The author mentions this, and said he tested for "alpha decay" by applying method to datasets that preceded the data on which the model was tested/trained.
Some of the papers I've seen are ridiculously obviously over-fitted.
For example published in 2018, but "tested" on 3 months of 2010 prices of GBP/USD, USD/SEK and USD/THB. Quality forex data is so easy to get freely, that picking 3 months from 8 years ago on one major pair and two other random minor ones just stinks.
>Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected
including methods that use:
>News Text Mining. - This is where they'd use NLP on headlines or the body of news as a signal.
I have to call this out.
Is this author suggesting that you couldn't have made money by shorting Enron stocks milliseconds after the scandal was made public? Is it impossible to make money by buying a stock in a small company, seconds after an acquisition is announced? If a CEO gets sent to prison, will that company's stocks not be affected?
And then there are other methods that use:
>Fundamental data. So ratios from the income statement/balance sheet
So buying stocks in companies with good financial health is not profitable?
they're probably not accounting for HFT. that's not unreasonable to expect.
> So buying stocks in companies with good financial health is not profitable?
everybody has the same common sense, prices reflect all available information (at least, if you believe the efficient market hypothesis, which I do to some extent). so you shouldn't expect the method to be profitable in excess of the overall market profit -- what we call alpha.
The market is not perfectly efficient. Small cap stocks outperform large cap stocks; the S&P500 outperforms leaving money under your mattress.
I am more than willing to bet that if you combine these methods with an algorithm that estimates a stock's "proper price" with the information, a sophisticated algorithm should be able to at least outperform a layman's "Buy-and-Hold" strategy.
You can put your money on that, but you would pretty consistently, it turns out, be wrong. Which you would know if you had read the OP.
Profiting off of other people's tendency to trade too much and too confidently is some of the surest money in the market, because regardless of evidence people want to believe that they can positively effect the outcome. Nest eggs are more like soufflés than caramels.
Published academic studies on the size factor go back decades. I'm not ready to consider them refuted because someone on reddit says he google searched papers from the last eight years and found them lacking, especially since he doesn't specifically claim to debunk any of the major factors, and the momentum factor he even confirms.
>Profiting off of other people's tendency to trade too much and too confidently is some of the surest money in the market
This proves my point; this statement is counter to the efficient market hypothesis, and it shouldn't be too difficult to algorithmically find trigger events that cause people to trade too much and too confidently.
> the S&P500 outperforms leaving money under your mattress
This doesn't conflict with the efficient market hypothesis. The S&P500 also outperforms 'investing' in blackjack. No one is claiming that holding your money as cash is on the efficient frontier.
>No one is claiming that holding your money as cash is on the efficient frontier.
But it would be in an efficient market. As more investors invest in more profitable assets, the price of those assets rise, which makes the return on those assets fall relative to the initial cost. The Efficient Market Hypothesis, in it's strongest form, implies that every asset is on the efficient frontier.
1) You‘d still get a risk premium, because only known information can be priced into the stock. Unkown information is risk.
2) Capital is rare. There is no unlimited supply, so it‘s distributed between assets as well as available information allows. But there is still unsatisfied capital needs where the money can be employed more efficient than holding it cash.
So an investor that can anticipate an increase in, decrease in, or general level of a) market risk, b) a market's risk premium, or c) available market capital, can predict market movements. Just because CFAs use fancy names for market imperfections doesn't mean that they're not exceptions to the EMH.
Yes, and value stocks outperform growth stocks over long enough periods of time too. But what does that have to do with the market not being efficient?
The academic explanation for why small cap and value outperform large cap and growth is that small cap and value companies are riskier investments. The factor risk premiums exist because investors need a higher return to reward them taking on more risk.
Risk premiums actually support the existence of an efficient market.
> Is this author suggesting that you couldn't have made money by shorting Enron stocks milliseconds after the scandal was made public?
Enron's collapse was 18 years ago. I suspect if this happened today, with today's trading environment, the answer to your question would be "yes." The algos today will parse an article, enter & exit a trade faster, than a human can read the headline.
> So buying stocks in companies with good financial health is not profitable?
That alone, probably not. You need to have an edge. If everyone else knows it's financial health clearly, then the price is already "bought up."
> That alone, probably not. You need to have an edge. If everyone else knows it's financial health clearly, then the price is already "bought up."
This is untrue. FAANG (all healtly companies) have been outperforming the S&P500 consistently. In fact, investing in FAANG is probably the "dumbest" smart play you can make. And, alas, you still come out on top.
Yes, and Enron was incredibly healthy too. Netflix has recently seen a downturn. The "health" of companies is not a constant. Can you predict when it will sour? Or are you certain that these companies will never fail? If so, why?
I think it's a bit of a stretch to say that FAANG stocks have "consistently" beaten the S&P 500. Most of the FAANG companies haven't even existed for long enough to draw meaningful conclusions from. The one that has (Apple) once underperformed the S&P 500 for 11 years from 1993 to 2004.
This brings an interesting point regarding reproducibility in economics. It's possible for a paper to be legit and be good science, but the moment it's published it becomes irreproducible because other actors are going to use the published approach from now on and the balance in a game theoretic way is not the same. By publishing a paper you can change the thing you are studying.
This is what the author is talking about when he says "alpha decay." He tried to account for it with backtesting (so simulating a market that has no knowledge of these strategies) and the strategies still failed.
The claim is that you can't do it consistently. Your sentiment detector has to more accurately capture the state of a randomly selected set of companies (not one selected with the benefit of hindsight, like Enron) based on news sentiment than the information already incorporated into the stock's price.
The important concept to understand about profitable investing is that you have to have a strategy that others are not also using.
Sure, investing in companies in good financial health is profitable. Unless everyone else does it too, and they drive up the price of the profitable companies, until all upside is gone (i.e., price is baked in). You're not better at finding profitable companies than anyone else.
Shorting stock on headlines? Sure, if you can beat everyone else. (You can't.)
The other is merely stating that, according to his analysis, apparently all these strategies did not bring an edge to the market.
"The important concept to understand about profitable investing is that you have to have a strategy that others are not also using."
People always say this, but it doesn't make sense to me.
If I go to the grocery store and buy produce, and I assume I'm not more knowledgeable than an expert purchaser for a food service business, then am I necessarily better off buying stuff at random without even looking at it? If I'm really clueless, can I not learn to choose good stuff by examining it and trying it, and finding out what other people look for?
The problem, I think, is that when people are looking for a way to beat the market, they, pretty much without exception, look for ways to process the predigested information about a select group of companies that is already in a structured form. That stuff is the information that's most absorbed into prices, so why not stop "looking for your keys under the streetlight"?
Precisely because the market is very efficient at pricing everything that people can quantify, you don't have to quantify much at all! If you can tell rotten fruit from fresh, then you are adding value and you can assume that everything that you find difficult to evaluate is already factored in.
What if you just spent 10 seconds looking at each business description of a public company's 10-K? Instead of looking at numbers at all, just treat your investing like you have a stack of several thousand resumes and you need to hire 20. Of course, you want to look at some numbers later on before buying, just like doing a background check on a prospective employee, but just looking at a broad cross section of what's out there is enough to observe really obvious and educational patterns.
Here's something I read in a 10-K recently:
"In July 2014 as part of a diversification strategy we acquired companies engaged in the manufacture and marketing of electro-hydraulic servo-valves and the development of optical fiber hardware and software solutions for the security and protection industry. However, in the fourth quarter of 2015, we decided to focus on our hog farming operations, sell our operations in electro-hydraulic servo-valves and optical fiber based security & protection and seek to grow through internal expansion and acquisitions of businesses in the agricultural industry."
Now, if the market is efficient, and I am not an expert in hog farming, servo valves, or investing then how can I tell if this company is a better or worse bet (at the current price) than, say, Apple?
I see where you're going with the grocery store produce analogy, but it's not a useful analogy for how markets work.
1. Those grocery store items (let's say tomatoes) are priced equally to each other, by decision of the supermarket. This is not true of securities, which are individually traded and priced separately, which allow differences to be priced in.
2. The market for tomatoes at your local grocery store is geographically constrained, limiting the number of participants. This is not true of publicly traded stocks that nearly anyone in the world can trade.
3. The amount of money that can be put to use finding efficiencies at your local market is small. In global markets, the payoff is in the billions, and so many people are scouring for similar efficiencies - and in the process, eliminating them.
4. The tomato market has high transaction costs. What are you going to do when you find a better tomato for the same price? Sell it to someone else for a higher price? No. The arbitrage opportunity for tomatoes doesn't exist.
Efficient markets is not a thesis that is required to be true of all markets by some sort of law. They are a consequence of liquid markets that are large and traded by many well-funded participants. The analogy must have the same factors to be valid.
Stocks do not tend to go up, that would imply a greater than 0 return average for series. We instead get mean blur and skewness (which is actually often to the left).
The aggregate of the traded stocks, i.e. the market, goes up on average. That's why you make money by holding diversified portfolios.
> Is this author suggesting that you couldn't have made money by shorting Enron stocks milliseconds after the scandal was made public?
No, he's claiming that any of the actual published systems, if they would correctly have made money on that one special event, would not do so on enough other events to make up for those they would lose money on plus transaction costs on all the trades they would make to actually beat a broad market index.
It's pretty easy to (with hindsight) design a system that would make money on Enron or any other isolated event. It's harder to build a system that will consistently beat the market on future events that it's not designed against.
The stock market is a complex adaptive system where the agents are constantly changing their strategies so that even if you were to find inefficiencies or patterns, they are only ephemeral.
> The easiest way to test whether it was truly Alpha decay or just overfitting by the authors is just to reproduce the paper then go further back in time instead of further forwards. For the papers that I could reproduce, all of them failed regardless of whether you go back or forwards. :)
This is why the truly successful quant groups like Renaissance continuously adjust their strategies and come up with new ones. Renaissance in particular has invested heavily into their data processing pipeline which enables them to have a significant advantage over the rest of the field.
Yes, if I remember correctly from an interview I saw with the founder he mentioned that the barrier to entry used to be high and that commodities markets 'used to' trend.
I had the same result from most papers. My personal conclusion is that when people get a winning strategy they don’t publish. I personally put my money where my mouth was for a few years:
That being said, I try to be honest too. This can disappear any time and the model I use may only be good in this environment. I do not know. I think that’s the challenge with papers, you don’t honestly know when or if the strategy works. It clearly won’t forever regardless.
That’s why I don’t share my exact method. And after doing all the research myself AND trying to sell my algorithm. I honestly don’t think the industry knows what it’s doing either. People are worried about sharpe ratios and all this BS stuff. The reality is for these models you mitigate risk via temporary and ever changing methods. Can’t really publish on that.
I hope that your personal implementation of your strategy takes into account 2008 ;). Because, hooo, boy, your 20% YoY returns sound great, but may not be taking enough risk of a similar collapse into account.
It does take those risks into account and I tested it on 2006-2008 previously. However, there is little point in testing a collapse with the same model. If the stock market collapses anyone investing would be out of luck. Your standard models won’t be able to track that, because it’s usually something like a “black swan” event[1].
Instead you’d want some sort of meta model. In either case. I’m getting 100% YoY returns when I augment my model, in real life. I think even a 50% loss one year wouldn’t be the end of the world.
Winning strategies are not published generally speaking. You wont see many business people blogging in detail about how they went for 0 to 1. Which is shy most blogs about business are a load of BS. Reading your statement and this article confirms my long suspicion that relying on internet strategies is a bad startegy.
There are definitely winning strategies. The issue is whether they can be reliably identified beforehand (thus rewarding skill), or whether people who implement them are just lucky (regardless of whether they believe they were skilled or not).
Well, if you get 30 persons to flip a coin six time, there is good chance one of them will get all tails or all head. Now ask him his strategy for such amazing coin flipping skills! (example taken from the book "statistics done wrong"[1]).
I would be interested in seeing a total distribution of hedge funds return, not just outliers.
That would be the reversion to the mean[1]. This is a term I really dislike because it makes it sounds as if there is some sort of equalizing force making over-performers under-perform later. This is more the following: if you overestimates the expectation of a random process, you are going to be disappointed.
In our case, this does not makes the "hot streak" any less probable when you start looking, for a specific edge fund. It is true it would be interesting to select a group of over-perfomer and study their future return to know if past performances are a good predictor of future performance. I feel you would probably get mixed results!
Although as I said in the sibling comment, you are probably right, and no amount of statistics could explain performances seen in this particular edge fund.
I understand what you are saying, but the equivalent to the Medallion fund would be something more akin to winning that same coin flip 30 times in a row. They have been running approximately 30 years, beating market returns by a significant margin, year after year. The 40% average is net of fees, so the overall return of their strategy has actually been higher than that. The odds that they have accomplished these returns through luck alone are astronomically low.
This is true, though I don't know how many edge funds operate. I guess some people really have working strategy!
The example I gave was more of a word of caution regarding past performances as indicator of expertise, but I guess I made it sound more generalizable than needed!
The trick is this: the chances that a specific fund will do well that long via luck are very low, but the chances that there exists a fund among all that exist that has done well via luck are quite high.
I can tell you haven't actually done the calculation. There's only been about 20000 hedge funds in total over history. The odds of random chance producing Medallion's track record with that many draws are actually very low.
It's probably still the Medallion Fund. Except it doesn't take outside money.
One characteristic of a good money manager is that he knows when to start rejecting money. Large AUMs tend to converge by necessity towards index funds (or something underperforming an index fund).
Is your "winning" strategy so different from other social data / sentiment analysis approaches? There is some novelty with how you weigh the sentiments (based on how much of an insider or expert they are) but I am sure existing trading strategies weren't just taking a dumb average of a twitter firehose either. Shouldn't it be easy for some large firm to replicate your approach and make the alpha disappear?
> Shouldn't it be easy for some large firm to replicate your approach and make the alpha disappear
First, I don't think alpha ever fully disappears.
Second, after speaking with twenty or so firms very few are using sentiment directly. Those that do, I suspect don't take the additional steps to build complex NLP based systems and weight insiders/experts. Even if you weight experts, the methods for doing so are also complicated (cross check against LinkedIn should be easy enough, but also limits information).
Anyway, I personally haven't seen much difference in the back testing.
Just a guess, but: if you come up with a successfull algorithm, you still need to have money that can be invested in order to use the algorithm. So maybe someone else with 100x more money to invest would pay more for the algorithm than you could earn from it in a lifetime.
> I had the same result from most papers. My personal conclusion is that when people get a winning strategy they don’t publish.
This is kind of obvious to me. It is also the reason the OP posted their results. I'm sure if they found one strategy that worked, after putting that much time into their research it would be really stupid to announce to the whole world that it works.
On the other hand, in the trading world where everyone is a competitor, you might want to deliberately introduce some confusion - but it looks like plenty of actors are doing this anyway.
It seems that you consider the sharpe ratio to be not worthwhile. Would you be able to elaborate on that point? As someone who is getting into algotrading, I’m currently using the sharpe ratio to quantify risk, but would like to hear another take on it.
Back in the 90's I traded with TradeStation, yes, always running my various algo's on "historical data" - - - which was ALWAYS just curve fitting.
I once had formula that made ton's of $$$$ on 4 years of backtested historical data, S&P futures, but once I started paper trading it it failed within one month.
Has anyone tried a simple approach - trying to predict which of the S&P 500 will have the lowest 10% returns, and build an (S&P 500 - 10%) index? It seems obvious that the S&P is stacked with some great companies and some old dogs. Does that method not work?
Since it's obvious, everyone knows it. And since everyone knows it, it's already priced in. You cannot find an edge by acting on widely-known public information.
I’m acting on the fact that almost all my investments are in the S&P 500, and I don’t have a (easy) way to pick my favorite 450 out of those 500. How would knowing the worst 50 be already priced in? They would be priced lower than they should? Good let’s get them out of my index. How else?
They'd be priced low. There is no "should" when it comes to markets - the market price is whatever people will transact at.
The problem is that stock price movements depend upon future events - making money in the markets is effectively a future-prediction problem. So if your strategy is to discard the bottom 10%, great, you got rid of Foot Locker and Sears. You also would've gotten rid of Apple in 1998, which was responsible for a good portion of the index's gains over the last 20 years. And you would've kept losers like PG&E, which went bankrupt over a black-swan event (they were doing fine until they burned down a town).
Prices of stock are based on future expectation of earnings. This is why you can have a company like Amazon have incredible earnings, but still have the share price tank. IE, expectations were that they would have even more incredible earnings, but they were merely incredible.
This is also how you can have companies that are on the verge of bankruptcy get huge stock gains if they defy earnings expectations.
As such, expectations of future results are already priced into stocks. It's literally the stock price. Whether a stock moves up or down is based on the delta between reality and expectation.
There are no quick easy hacks that give you reliable above market returns. If there were, enough people would use them that the pricing would correct for it because of demand and the above market return opportunity would disappear.
Intuitively the problem you run into is that occasionally those obviously bad companies have amazing comebacks and you completely miss out on those, so you can still end up underperforming the S&P 500.
If you know that 5.6% of the users (automated, or parroting the preacher) of an exchange will use fibonacci retracements, and that 15% of the amateur market will follow the market price change caused by either buying or selling activity, then you can play roulette with a decent edge. Of course, not as much as the preacher, who is allowed to bet before (s)he will speak to his or her congregation.
When you gather enough of these commonly used technical analysis, it's like having to predict in which startup Ron Conway will invest, but you can calculate Conway in a Python one-liner, and keep up-to-date by going to weekly sermon.
yes, this is possible, with more trades becoming a self fulfilling prophecy, and have stop limits in place all the times it didn't work.
I wish the technical analysis flock would merely incorporate different things. TA takes a time series chart and imagines time series patterns. It primarily neglects what didn't get printed on a time series chart, and why. How big are the orders at the resistance level, do you have a record of the order sizes that appeared at the last resistance level? who is selling at the resistance level and why? The TA answer is "just because its a psychologically round number for the resistance price" or "because thats how high the last high candle was", but you can greatly improve your win rate by understanding who is in the market and why, which is possible to understand and a large portion of my trading strategies. It can be much more data intensive though so I can see why 1980s gurus did not do it.
I could bet a ton that most people will make excuses as to why the papers failed. There's something within us that wants to hit the stock market lottery.
I truly believe that there are streeks to profits in the stock market in the same way you will find streeks in any set of random numbers but they are impossible to find in a consistent manner.
The road to wealth for most in the stock market is time and investing in a basket of good stocks.
Whoever thinks that they have found a system to profits in the stock market. Test and retest your method a few times. It's unlikely you have a winning system.
Describe one. I know if there are any current ones no one will talk about them. But have there been any consistent ones in the last 100 years? I'm sure someone would have at least written about them in their memoirs.
I don’t know any good ones... AFAIK value, size signals, merger arbitrage, and other kinds of predictionsused to work well in context of statistical arbitrage... but yeah their existence is evidenced by successful companies / hedge funds (Renaissance, Citadel, Jump, Two Sigma, ...)
But there is a foolproof way of profiting from the stock market. Insider trading! I find it fascinating that people do not believe it occurs at a grand scale given the low risks and huge rewards. Exactly like how people believe athletes don't use steroids so they get all upset when every once in a while one is caught. :)
The first hedge fund Jones & Co. that launched 1949 hired business writers that would consistently "find" profitable trades. The logic was that Jones & Co. assumed that they had a line to insider information.
Yup, you can make a consistent buck if you know the right people with the right info. That system will always work.
Hasn't there recently been shown that only strategies using simple momentum-derived technical indicators were able to consistently bring returns in the stock market?
Not just recently, it's been known for a while. The author actually confirmed it:
> Almost every instrument is mean-reverting on short timelines and trending on longer timelines. This has held true across most of the data that I tested.
(But momentum isn't the only thing that appears to work.)
Could you expand on that? I've only skimmed the article, but I don't see any "crypto scams" pushed in the article. The article is just about something the author seems to know about (algorithmic trading), applied to the cryptocurrency market. It does promote the authors project (why else would you write a Medium post), but in the worst case, that would be a normal scam and not a crypto scam.
The whole article just seems like an attempt to steal money from uninformed people. He starts by giving vague information about trading strategies in general, then linking to an article about Renaissance Technologies as an example for successful algorithmic trading, then stating that most trading bots aren't successful and that the crucial differentiator for deciding whose bots to trust is the person's professional experience, which is obviously a reasonable thing to do, however the picture at the beginning of the article of him at a trading desk and him repeatedly mentioning his 7 year experience as a trader combined with the complete lack of any actual proof that his bots are actually profitable, make it seem as if he just tries to profit off his previous experience.
He ends the article writing this:
"All you’ll need to get started is:
1. $1000
2. To press a single button to get the bots started"
Furthermore in the reddit comments in response to the following question: "130 papers re-implemented in 7 months? I'm blown away. Write a software engineering book about how you did it so quickly. Then write a self-help book about having enough motivation to see it through." he writes:
"100-hour weeks and a desire for a better life for the ones you love will get you there pretty quickly"
This guy seems like a complete fraud, I find it sort of sad that this has landed on the frontpage of HN with that money upvotes.
Oh don't get me wrong, I wouldn't trust the guy farther than I can throw him either. I just wanted to be a bit pedantic about "normal fraud" vs. "crypto fraud" (= ICO or similar).
There was a fair amount of overlap in the papers. Makes the testing much easier.
> So with the papers, I found as many as I could, then I read through them and put them in categories and then tested each category at a time because a lot of papers were kinda saying the same things.
more important than this study if done correctly is the fact that he built a framework that can ingest all this data and that he had access to all these datasets
typical hedge fund spends millions of dollars in order to build such frameworks and buying datasets, sure most academic papers fail if you replicate but the framework and datasets are very valuable because you can eventually find something on your own or an improvement on existing ideas if you keep trying hard enough + there are other sources of ideas like quant research from brokers, ideas from platforms like quantopian etc. but yea in general if you have an outstanding idea that works - you would have very less or no incentive to publish it. why would jim simons have his researchers publish anything when they can make money for him all day long everyday ... just my 2 cents.
i asked the author where did he source his data, his reply was "i scraped data"
how can you scrape pricing data? not every data in this is on public domain, otherwise there would be no Bloombergs, CapitalIQs selling data for millions(sure they're overrated and overpriced but still!). Or in other words, if he is right - he can sell data and make millions. no need of looking for an investment strategy. just my skeptical side saying :)
you need clean data to accurately test ideas. for instance getting tick data is quite expensive. most universities have free access to Bloomberg, CapitalIQ etc. datasets the reason professors can test and also the reason some smart guys in the industry work for university on the side
it's not enough data to test all of those 130 papers, he said he has also tested short term reversion trends. for that you need tick data(or atleast minute/hourly data)
Agreed — something about this post doesn’t pass the smell test.
1. He’s a profitable trader at a tier 1 firm who has the spare time to not only develop a series of algorithms based on 130 research papers, but also sufficiently backtest them in 7 months?
2. He said he looked at the past 8 years of papers, but refers to multiple models correctly predicting the 2008 financial crisis.
3. Where are the code samples?
Edit:
Lol, just realized his medium post ends with a crypto scam.
Ditto on the "130 papers in 7 months." I am not familiar with the field, but I assume the process would look like this:
* Read and understand paper
* Find and download appropriate input data
* Code paper model and validate (he said he wrote his own code)
I can see myself being able to do this for ONE paper in maybe a week. He claims he was doing 1-2 of these per day. Wow. So either there is some exaggeration on his part, or he is a total wizard in his field.
I think you are a quick study; typically it takes me a week to figure out the detail of what's being done in a paper. Getting the data and testing would take longer. Charitably he/she has a framework with all the data required sitting ready to go and is just writing wrappers to the models. But even downloading frameworks from Github and getting them working takes a couple of days - for me. For example I've been playing with the Graph-network code from deepmind for a few weeks - I had to learn how the graphs were represented, how to build them and access them and how the models were made and put together. Just working that out was a solid three day job. Now I can build things and test out what's going on in the examples and get a feel for the framework, probably (if there was a problem) I would be in a reasonable position to say "this doesn't work like they think it does" (it does, but no surprise) but unless you've done that leg work I think you can't really. I think a proper replication effort is really 1 man month of expert time - or really you're just throwing stones.
Depending on the complexity of the model it would take me at least a month for a single paper. What makes it fully unbelievable to me is the claim of detecting p-value hacking in many of these 130 papers while doing 3 papers every 2 days.
To make that claim for a single paper I would 1. have to be able to reproduce their p-value, and 2. spend enough time with the model to understand how/what assumptions were unfairly tweaked to get to that p-value.
Just running your own implementation of a model on your own dataset and getting an insignificant or different p-value is not enough. You might just have implemented the model wrongly.
the author is trying to sell crypto trading bot ! looks shady to me
if anything guarantees you profit, run away from it as fast as you can :
https://credium.io/
I have true hate for the authors of this paper: "A deep learning framework for financial time series using stacked autoencoders and long-short term memory". Probably the most complex AND vague in terms of methodology and after weeks trying to reproduce their results (and failing) I figured out that they were leaking future data into their training set (this also happens more than you'd think)."
- Not sure how author tried to implement it , but is this not how you train LSTM networks by feeding t+1 data back into the cell again to predict t+2 data. It will be easier if author made it open source as well
Leaking future data in would be using t+1 for t, e.g. something like a bi-directional LSTM. I assume he means the actual training dataset had some kind of signal in the data that was also in the test data.
People do this by doing things like testing that their features contain information in both the training and test set. Because they are not exposing the data directly to the classifier they think that they haven't compromised the test set - but what they have done is increased the chances of a chance correlation.
I wrote my thesis last year comparing different RNNs against each other using this exact paper as baseline and basically concluded that you would be better off predicting the price yesterday than using their results. Authors did not respond when prompted for implementation details or comments.
Overall, concluded that amongst RNNs the GRU architecture proved most favorable but still would not outperform simple stochastic models of the financial industry toolbox.
Best survey of the subject I have found (most are bullshit) is Finding Alpha by Eric Falkenstein which he has graciously offered for free on his blog. By the subject I mean, finding an edge in trading. Spoiler alert there is no system to follow. He wrote algos for a local options market maker, had a significant Econ PHD thesis. The basic premise of the book is that real Alpha is rare, idiosyncratic and gets exploited by those in the know and eventually the edge goes away after several years. He gives several classic historic examples which are what made this book so interesting and unique to me.
http://falkenblog.blogspot.com/2016/08/finding-alpha-pdf.htm...
214 comments
[ 3.6 ms ] story [ 245 ms ] thread> Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected (I guess ultimately they're all the same thing but still) OR a few may have had a smidge of Alpha but as soon as you add transaction costs it all disappears.
> I should caveat that I was a profitable trader at multiple Tier-1 US banks so I can say with confidence that I made a decent attempt of building whatever the author was trying to get at.
> Almost every instrument is mean-reverting on short timelines and trending on longer timelines.
i.e. he confirms the momentum factor, which isn't surprising since there's more solid evidence for it than anything else, going back hundreds of years.
He doesn't say what fundamental factors he looked at, so it's possible that value, size, and profitability/quality would hold up as well. All those have been studied pretty extensively in academia, in papers going back decades. The author took only a fairly random sampling of recent papers.
Not exactly sure but I think the "fundamentals" are top-line revenue, unit cost, unit margin, yoy-growth, EBTDA, cash-on-hand, default-alive - and likely the ratios those values produce.
Two simple factors that are well supported: on a risk-adjusted basis, stocks of small companies do better than big ones, and value stocks do better than growth stocks. ("Value" means the stock is cheap relative to some simple fundamental measure like the company's book value.)
A simpler way to think of it is any data that comes off the Income Statement, Balance Sheet, or Cash Flow Statement - or can be directly calculated using such data.
Having a small impact is not the same a having zero impact.
Traders operate over longer timeframe even if they are holding an individual stock for minutes there is a limited pool of stocks. Keep playing the game and longer term impacts add up.
Look, I know writing stuff up sucks. But it could be a great opportunity to learn a lot of things from a very knowledgable person. With the right prof, it could be a great undergrad project.
Except plenty useful came up. "Literally every single paper was either p-hacked, overfit, or a subsample of favourable data was selected" out of 130+ is a significant result. The media would jump on this, and they just might even without any proof of work.
If it worked, it wouldn't be published, or at least not until it stopped working.
Except they didn't share the results with the rest of us. I know you said "experience" not "results", but when disproving papers, the least you can do is write down three sentences about each paper as you go along reproducing them, noting what you are seeing, perhaps with a snapshot (just a zip file or so) of the code. This is calling a whole field nonsense (that everyone expects to be full of nonsense) without giving enough evidence for anyone else to dispute your claims.
[0] https://www.scottaaronson.com/papers/pnp.pdf
[1] https://arxiv.org/abs/1002.2284
Edit for downvoters and repliers: If enough market participants are irrational, then it can still be possible for people to predict other people, instead of the market, and make money that way.
NP-hardness indeed doesn't rule out heuristic approaches, but experience with 3-SAT and other NP-complete problems suggest that there will be arbitrarily bad times, and that in those times, the amount of loss can be exponential in the length of time that the heuristic poorly predicts the market.
Half the battle is not doing something stupid.
According to the paper, if you don't believe that P=NP then you believe that the market is inefficient, which means there's profit to be made. The paper even suggests how.
I could have told you that without testing. If anyone had a lucrative strategy would they disclose it in a paper to the general public? I think not.
If the guy is lying and it's all made up, I'd still wager that if you assembled a collection of 130 papers on this topic, at most 10% would be valid [1]. So even if he's wrong, he's probably not wrong.
[1] I have an academic background, and I think invalid papers make it through peer review constantly. Peer review is not what people think it is.
I would even guess that a paper being published means that, at the point the paper started to be written, its alpha had already decreased to zero. Otherwise the writers of the paper would still be using that approach. That's how it can appear to provide no value even if you extrapolate it back in time.
For example published in 2018, but "tested" on 3 months of 2010 prices of GBP/USD, USD/SEK and USD/THB. Quality forex data is so easy to get freely, that picking 3 months from 8 years ago on one major pair and two other random minor ones just stinks.
including methods that use:
>News Text Mining. - This is where they'd use NLP on headlines or the body of news as a signal.
I have to call this out.
Is this author suggesting that you couldn't have made money by shorting Enron stocks milliseconds after the scandal was made public? Is it impossible to make money by buying a stock in a small company, seconds after an acquisition is announced? If a CEO gets sent to prison, will that company's stocks not be affected?
And then there are other methods that use:
>Fundamental data. So ratios from the income statement/balance sheet
So buying stocks in companies with good financial health is not profitable?
Something's being left out here.
> So buying stocks in companies with good financial health is not profitable?
everybody has the same common sense, prices reflect all available information (at least, if you believe the efficient market hypothesis, which I do to some extent). so you shouldn't expect the method to be profitable in excess of the overall market profit -- what we call alpha.
I am more than willing to bet that if you combine these methods with an algorithm that estimates a stock's "proper price" with the information, a sophisticated algorithm should be able to at least outperform a layman's "Buy-and-Hold" strategy.
Profiting off of other people's tendency to trade too much and too confidently is some of the surest money in the market, because regardless of evidence people want to believe that they can positively effect the outcome. Nest eggs are more like soufflés than caramels.
This proves my point; this statement is counter to the efficient market hypothesis, and it shouldn't be too difficult to algorithmically find trigger events that cause people to trade too much and too confidently.
This doesn't conflict with the efficient market hypothesis. The S&P500 also outperforms 'investing' in blackjack. No one is claiming that holding your money as cash is on the efficient frontier.
But it would be in an efficient market. As more investors invest in more profitable assets, the price of those assets rise, which makes the return on those assets fall relative to the initial cost. The Efficient Market Hypothesis, in it's strongest form, implies that every asset is on the efficient frontier.
1) You‘d still get a risk premium, because only known information can be priced into the stock. Unkown information is risk.
2) Capital is rare. There is no unlimited supply, so it‘s distributed between assets as well as available information allows. But there is still unsatisfied capital needs where the money can be employed more efficient than holding it cash.
Yes, and value stocks outperform growth stocks over long enough periods of time too. But what does that have to do with the market not being efficient?
The academic explanation for why small cap and value outperform large cap and growth is that small cap and value companies are riskier investments. The factor risk premiums exist because investors need a higher return to reward them taking on more risk.
Risk premiums actually support the existence of an efficient market.
Additional reading:
- https://faculty.chicagobooth.edu/john.cochrane/research/pape...
- https://www.investopedia.com/ask/answers/022715/are-small-ca...
- https://www.investopedia.com/terms/v/valuestock.asp
(Strong form market efficiency has been disproven already, so this is weak form efficiency).
That's enough to make me believe the efficient market hypothesis is BS.
Enron's collapse was 18 years ago. I suspect if this happened today, with today's trading environment, the answer to your question would be "yes." The algos today will parse an article, enter & exit a trade faster, than a human can read the headline.
> So buying stocks in companies with good financial health is not profitable?
That alone, probably not. You need to have an edge. If everyone else knows it's financial health clearly, then the price is already "bought up."
This is untrue. FAANG (all healtly companies) have been outperforming the S&P500 consistently. In fact, investing in FAANG is probably the "dumbest" smart play you can make. And, alas, you still come out on top.
A sound strategy would also hedge against black swan events — so some money might be in gold or jewels or something. But that’s beside the point.
Sure, investing in companies in good financial health is profitable. Unless everyone else does it too, and they drive up the price of the profitable companies, until all upside is gone (i.e., price is baked in). You're not better at finding profitable companies than anyone else.
Shorting stock on headlines? Sure, if you can beat everyone else. (You can't.)
The other is merely stating that, according to his analysis, apparently all these strategies did not bring an edge to the market.
People always say this, but it doesn't make sense to me.
If I go to the grocery store and buy produce, and I assume I'm not more knowledgeable than an expert purchaser for a food service business, then am I necessarily better off buying stuff at random without even looking at it? If I'm really clueless, can I not learn to choose good stuff by examining it and trying it, and finding out what other people look for?
The problem, I think, is that when people are looking for a way to beat the market, they, pretty much without exception, look for ways to process the predigested information about a select group of companies that is already in a structured form. That stuff is the information that's most absorbed into prices, so why not stop "looking for your keys under the streetlight"?
Precisely because the market is very efficient at pricing everything that people can quantify, you don't have to quantify much at all! If you can tell rotten fruit from fresh, then you are adding value and you can assume that everything that you find difficult to evaluate is already factored in.
What if you just spent 10 seconds looking at each business description of a public company's 10-K? Instead of looking at numbers at all, just treat your investing like you have a stack of several thousand resumes and you need to hire 20. Of course, you want to look at some numbers later on before buying, just like doing a background check on a prospective employee, but just looking at a broad cross section of what's out there is enough to observe really obvious and educational patterns.
Here's something I read in a 10-K recently:
"In July 2014 as part of a diversification strategy we acquired companies engaged in the manufacture and marketing of electro-hydraulic servo-valves and the development of optical fiber hardware and software solutions for the security and protection industry. However, in the fourth quarter of 2015, we decided to focus on our hog farming operations, sell our operations in electro-hydraulic servo-valves and optical fiber based security & protection and seek to grow through internal expansion and acquisitions of businesses in the agricultural industry."
Now, if the market is efficient, and I am not an expert in hog farming, servo valves, or investing then how can I tell if this company is a better or worse bet (at the current price) than, say, Apple?
1. Those grocery store items (let's say tomatoes) are priced equally to each other, by decision of the supermarket. This is not true of securities, which are individually traded and priced separately, which allow differences to be priced in.
2. The market for tomatoes at your local grocery store is geographically constrained, limiting the number of participants. This is not true of publicly traded stocks that nearly anyone in the world can trade.
3. The amount of money that can be put to use finding efficiencies at your local market is small. In global markets, the payoff is in the billions, and so many people are scouring for similar efficiencies - and in the process, eliminating them.
4. The tomato market has high transaction costs. What are you going to do when you find a better tomato for the same price? Sell it to someone else for a higher price? No. The arbitrage opportunity for tomatoes doesn't exist.
Efficient markets is not a thesis that is required to be true of all markets by some sort of law. They are a consequence of liquid markets that are large and traded by many well-funded participants. The analogy must have the same factors to be valid.
Not quite, because stocks tend to go up.
What really is hard is making more profit than just holding stocks would. Because that takes actual new ideas.
The aggregate of the traded stocks, i.e. the market, goes up on average. That's why you make money by holding diversified portfolios.
No, he's claiming that any of the actual published systems, if they would correctly have made money on that one special event, would not do so on enough other events to make up for those they would lose money on plus transaction costs on all the trades they would make to actually beat a broad market index.
It's pretty easy to (with hindsight) design a system that would make money on Enron or any other isolated event. It's harder to build a system that will consistently beat the market on future events that it's not designed against.
Market is irrational in the short term. Hindsight is 20:20.
> The easiest way to test whether it was truly Alpha decay or just overfitting by the authors is just to reproduce the paper then go further back in time instead of further forwards. For the papers that I could reproduce, all of them failed regardless of whether you go back or forwards. :)
https://austingwalters.com/backtesting-our-100-yoy-profit-ge...
That being said, I try to be honest too. This can disappear any time and the model I use may only be good in this environment. I do not know. I think that’s the challenge with papers, you don’t honestly know when or if the strategy works. It clearly won’t forever regardless.
That’s why I don’t share my exact method. And after doing all the research myself AND trying to sell my algorithm. I honestly don’t think the industry knows what it’s doing either. People are worried about sharpe ratios and all this BS stuff. The reality is for these models you mitigate risk via temporary and ever changing methods. Can’t really publish on that.
Instead you’d want some sort of meta model. In either case. I’m getting 100% YoY returns when I augment my model, in real life. I think even a 50% loss one year wouldn’t be the end of the world.
[1] https://en.m.wikipedia.org/wiki/Black_swan_theory
There are people who are addicted to tracking weather fluctuations on Pluto and there are people who are not.
https://www.bloomberg.com/news/articles/2019-03-07/jim-simon...
I would be interested in seeing a total distribution of hedge funds return, not just outliers.
[1]:https://www.statisticsdonewrong.com/
Is that what happened?
In our case, this does not makes the "hot streak" any less probable when you start looking, for a specific edge fund. It is true it would be interesting to select a group of over-perfomer and study their future return to know if past performances are a good predictor of future performance. I feel you would probably get mixed results!
Although as I said in the sibling comment, you are probably right, and no amount of statistics could explain performances seen in this particular edge fund.
[1]https://en.wikipedia.org/wiki/Regression_toward_the_mean
The example I gave was more of a word of caution regarding past performances as indicator of expertise, but I guess I made it sound more generalizable than needed!
One characteristic of a good money manager is that he knows when to start rejecting money. Large AUMs tend to converge by necessity towards index funds (or something underperforming an index fund).
First, I don't think alpha ever fully disappears.
Second, after speaking with twenty or so firms very few are using sentiment directly. Those that do, I suspect don't take the additional steps to build complex NLP based systems and weight insiders/experts. Even if you weight experts, the methods for doing so are also complicated (cross check against LinkedIn should be easy enough, but also limits information).
Anyway, I personally haven't seen much difference in the back testing.
* Don't know what trading strats trade at
This is kind of obvious to me. It is also the reason the OP posted their results. I'm sure if they found one strategy that worked, after putting that much time into their research it would be really stupid to announce to the whole world that it works.
On the other hand, in the trading world where everyone is a competitor, you might want to deliberately introduce some confusion - but it looks like plenty of actors are doing this anyway.
Back in the 90's I traded with TradeStation, yes, always running my various algo's on "historical data" - - - which was ALWAYS just curve fitting.
I once had formula that made ton's of $$$$ on 4 years of backtested historical data, S&P futures, but once I started paper trading it it failed within one month.
The problem is that stock price movements depend upon future events - making money in the markets is effectively a future-prediction problem. So if your strategy is to discard the bottom 10%, great, you got rid of Foot Locker and Sears. You also would've gotten rid of Apple in 1998, which was responsible for a good portion of the index's gains over the last 20 years. And you would've kept losers like PG&E, which went bankrupt over a black-swan event (they were doing fine until they burned down a town).
This is also how you can have companies that are on the verge of bankruptcy get huge stock gains if they defy earnings expectations.
As such, expectations of future results are already priced into stocks. It's literally the stock price. Whether a stock moves up or down is based on the delta between reality and expectation.
protip: a fibonacci retracement from a randomly selected extreme will always tell you something
protip: it takes 5 months for your congregation’s account to get eaten up from transaction costs when their stop limits keep getting hit
in the mean time you can just play TA roulette and they’ll always be impressed by your “uncanny” perceptive abilities
When you gather enough of these commonly used technical analysis, it's like having to predict in which startup Ron Conway will invest, but you can calculate Conway in a Python one-liner, and keep up-to-date by going to weekly sermon.
I wish the technical analysis flock would merely incorporate different things. TA takes a time series chart and imagines time series patterns. It primarily neglects what didn't get printed on a time series chart, and why. How big are the orders at the resistance level, do you have a record of the order sizes that appeared at the last resistance level? who is selling at the resistance level and why? The TA answer is "just because its a psychologically round number for the resistance price" or "because thats how high the last high candle was", but you can greatly improve your win rate by understanding who is in the market and why, which is possible to understand and a large portion of my trading strategies. It can be much more data intensive though so I can see why 1980s gurus did not do it.
I truly believe that there are streeks to profits in the stock market in the same way you will find streeks in any set of random numbers but they are impossible to find in a consistent manner. The road to wealth for most in the stock market is time and investing in a basket of good stocks.
Whoever thinks that they have found a system to profits in the stock market. Test and retest your method a few times. It's unlikely you have a winning system.
But obviously nobody is going to publish such strategies. You’re looking at a negative selection of papers.
Yup, you can make a consistent buck if you know the right people with the right info. That system will always work.
> Almost every instrument is mean-reverting on short timelines and trending on longer timelines. This has held true across most of the data that I tested.
(But momentum isn't the only thing that appears to work.)
His advice? "Nobody knows nothing." (Read Bogleheads.org to see how to leverage this into wealth.)
What was the setup, how did you set up a pipeline? Was it R or Python? What was the data source?
I am more surprised by your productivity than anything else.
I’d also love source code + data for this; without it, it’s a claim with nothing to back it, yet.
Could you expand on that? I've only skimmed the article, but I don't see any "crypto scams" pushed in the article. The article is just about something the author seems to know about (algorithmic trading), applied to the cryptocurrency market. It does promote the authors project (why else would you write a Medium post), but in the worst case, that would be a normal scam and not a crypto scam.
He ends the article writing this:
"All you’ll need to get started is: 1. $1000 2. To press a single button to get the bots started"
Furthermore in the reddit comments in response to the following question: "130 papers re-implemented in 7 months? I'm blown away. Write a software engineering book about how you did it so quickly. Then write a self-help book about having enough motivation to see it through." he writes:
"100-hour weeks and a desire for a better life for the ones you love will get you there pretty quickly"
This guy seems like a complete fraud, I find it sort of sad that this has landed on the frontpage of HN with that money upvotes.
> So with the papers, I found as many as I could, then I read through them and put them in categories and then tested each category at a time because a lot of papers were kinda saying the same things.
typical hedge fund spends millions of dollars in order to build such frameworks and buying datasets, sure most academic papers fail if you replicate but the framework and datasets are very valuable because you can eventually find something on your own or an improvement on existing ideas if you keep trying hard enough + there are other sources of ideas like quant research from brokers, ideas from platforms like quantopian etc. but yea in general if you have an outstanding idea that works - you would have very less or no incentive to publish it. why would jim simons have his researchers publish anything when they can make money for him all day long everyday ... just my 2 cents.
how can you scrape pricing data? not every data in this is on public domain, otherwise there would be no Bloombergs, CapitalIQs selling data for millions(sure they're overrated and overpriced but still!). Or in other words, if he is right - he can sell data and make millions. no need of looking for an investment strategy. just my skeptical side saying :)
you need clean data to accurately test ideas. for instance getting tick data is quite expensive. most universities have free access to Bloomberg, CapitalIQ etc. datasets the reason professors can test and also the reason some smart guys in the industry work for university on the side
https://www.alphavantage.co/
1. He’s a profitable trader at a tier 1 firm who has the spare time to not only develop a series of algorithms based on 130 research papers, but also sufficiently backtest them in 7 months?
2. He said he looked at the past 8 years of papers, but refers to multiple models correctly predicting the 2008 financial crisis.
3. Where are the code samples?
Edit:
Lol, just realized his medium post ends with a crypto scam.
* Read and understand paper
* Find and download appropriate input data
* Code paper model and validate (he said he wrote his own code)
I can see myself being able to do this for ONE paper in maybe a week. He claims he was doing 1-2 of these per day. Wow. So either there is some exaggeration on his part, or he is a total wizard in his field.
To make that claim for a single paper I would 1. have to be able to reproduce their p-value, and 2. spend enough time with the model to understand how/what assumptions were unfairly tweaked to get to that p-value.
Just running your own implementation of a model on your own dataset and getting an insignificant or different p-value is not enough. You might just have implemented the model wrongly.
https://towardsdatascience.com/crypto-trading-bots-a-helpful...
I have true hate for the authors of this paper: "A deep learning framework for financial time series using stacked autoencoders and long-short term memory". Probably the most complex AND vague in terms of methodology and after weeks trying to reproduce their results (and failing) I figured out that they were leaking future data into their training set (this also happens more than you'd think)."
- Not sure how author tried to implement it , but is this not how you train LSTM networks by feeding t+1 data back into the cell again to predict t+2 data. It will be easier if author made it open source as well
Overall, concluded that amongst RNNs the GRU architecture proved most favorable but still would not outperform simple stochastic models of the financial industry toolbox.
You can check it out here https://github.com/jensgrud/financial-forecasting-lstm/tree/...