From the article: "It works by analysing the ‘top lists’ of editorial pictures on stock image website Getty. Using machine learning, the algorithm produces a daily score based on the types of photos used in global news reports. Lead author Dr Angel Zhong said investors could use this information to better predict daily stock market returns, based on the mood of investors worldwide."
Not, "Angel Zhong Fund announces 142% gain for year."
Exactly. If you could accurately predict returns, you wouldn’t publish, you’d profit. Profiting from the strategy would be a much better form of proof than a paper.
Although.. maybe this is the secret sauce that powers Jim Simons’ Medallion Fund
It probably had implicit status as post hoc reasoning since the images include post closure photos for at least some markets since some news images come after closure. They'd have to work awfully hard to demonstrate better than random with only prior data.
A confounding problem is that news sources change stories and use a/b testing to pick story winners.
This could even be a meta analysis of emerging sentiment on delay, which means you'd do better timewise on the origin sentiment than the second differential. Except you can't access the same pool of opinion so at best, it's a clever trick to access emerging sentiment through reactive reasoning on news sites. Why is that better as a driving function to invest across the market than other sources?
> Exactly. If you could accurately predict returns, you wouldn’t publish, you’d profit
While it's sometimes fun to assume things, it gets hard when you talk about people. In this case, not everyone is driven by profits and money, so it's most likely incorrect to assume "They didn't use this way of gaining profit themselves, so it's probably not profitable".
Also -- wouldn't it be a safe assumption that photo editors are choosing the pictures that match the tone of the news they are reporting on? This seems like a roundabout way of saying that financial news seems to mostly accurately report the state of financial markets....
Sure. That would be why using this or a similar method as a synthesis or proxy to aid prediction could have value, right? An issue is that methods of money management one might use in conjunction with this type of system can themselves have determinism for total returns. Measuring effectiveness would be hard.
Assuming the news gets printed before market open (not really a safe assumption in the digital age), for this to have predictive power one would have to be looking at the intraday difference from market open to close, or the difference from previous close to today's open, not from previous close to today's close. It's not clear from the abstract/article which they looked at.
That’s OP’s point: if this were truly predictive then you would just go raise a few billion dollars and quietly implement this strategy…you wouldn’t write a paper about it.
The fact that we’re reading about a paper, and not a headline about a wildly successful systematic fund, suggests this isn’t all that valuable.
You're probably correct in this case, but not in general. The authors of [1] could have quietly run a money printer but elected to seek reforms instead.
I’m not sure that’s the case in all circumstances. There’s at least a few financial anomalies that could be taken advantage of, but human biases constrain their actual use.
One example is the low volatility anomaly, where low risk assets outperform the market, when time and again, fund managers struggle to do so without leverage. But fund managers generally won’t employ such a strategy because their investors don’t want to see themselves in low-risk assets when high-risk ones are going gangbusters because of FOMO.
Can you elaborate beyond your one word? I’m assuming you’re referring to their ETF offerings, which are designed to match their respective market (minus fees) which means they are designed to not beat their benchmarks
I didn't read the whole study and this is a pre-proof but I'm not very optimistic about claims like this. There are often many data biases that are hard to root out.
People often claim to "find" new market factors - but in reality they're just re-hashes of some other very well known and mathematically defined factor. I think in this case, its just called "momentum".
Sentiment analysis is like alchemy. People have searched for it forever in hopes of some sort of pot of gold. Turns out, sentiment is directly caused by the markets. If you're in the red you're unhappy... its that simple.
Then what happens? Your favorite WSJ reporter will write up some pessimistic article with a stock photo of some long-lost stock broker dinosaur with his face in his hands. Market might go down another few days... momentum.
I haven't read the whitepaper - its paywalled. But probably not worth a read.
No test? No validation whatsoever?
Ah, found this: “Consistent with behavioral models, Photo Pessimism predicts market return reversals and trading volume. The relation is strongest among stocks with high limits to arbitrage and during periods of elevated fear.”
Anybody who is a bit familiar with the stock market knows that highly arbitraged stocks tend to fluctuate greatly with bad news. The effect here is magnified so it’s easy to say there is a correlation. But by no way it can represent the market sentiment as a whole. And IMO it doesn’t speak for the value of the model.
31 comments
[ 3.2 ms ] story [ 70.9 ms ] threadExcellent! I am sure the research was paid for by the Australian taxpayers, but it's Elsevier who makes money from publication.
Not, "Angel Zhong Fund announces 142% gain for year."
Although.. maybe this is the secret sauce that powers Jim Simons’ Medallion Fund
A confounding problem is that news sources change stories and use a/b testing to pick story winners.
This could even be a meta analysis of emerging sentiment on delay, which means you'd do better timewise on the origin sentiment than the second differential. Except you can't access the same pool of opinion so at best, it's a clever trick to access emerging sentiment through reactive reasoning on news sites. Why is that better as a driving function to invest across the market than other sources?
While it's sometimes fun to assume things, it gets hard when you talk about people. In this case, not everyone is driven by profits and money, so it's most likely incorrect to assume "They didn't use this way of gaining profit themselves, so it's probably not profitable".
In fact, one might get even better results if one trains the model on the daily market charts newspapers publish such as this: https://thumbs.dreamstime.com/x/stock-market-newspaper-backg...
Oh, that's harder to get than the top-10 list at Getty, I see. Yeah, convenience sampling is king.
Is one of the authors "Not Sure!" by any chance?
So now this won't work.
The fact that we’re reading about a paper, and not a headline about a wildly successful systematic fund, suggests this isn’t all that valuable.
[1] https://online.wsj.com/public/resources/documents/1029sechft...
One example is the low volatility anomaly, where low risk assets outperform the market, when time and again, fund managers struggle to do so without leverage. But fund managers generally won’t employ such a strategy because their investors don’t want to see themselves in low-risk assets when high-risk ones are going gangbusters because of FOMO.
https://sci-hubtw.hkvisa.net/10.1016/j.jfineco.2021.06.002
Which implies the articles for that day are already published.
Which means this classifier basically "reads the market news correctly based on the pictures within".
Sentiment analysis is like alchemy. People have searched for it forever in hopes of some sort of pot of gold. Turns out, sentiment is directly caused by the markets. If you're in the red you're unhappy... its that simple.
Then what happens? Your favorite WSJ reporter will write up some pessimistic article with a stock photo of some long-lost stock broker dinosaur with his face in his hands. Market might go down another few days... momentum.
I haven't read the whitepaper - its paywalled. But probably not worth a read.
Anybody who is a bit familiar with the stock market knows that highly arbitraged stocks tend to fluctuate greatly with bad news. The effect here is magnified so it’s easy to say there is a correlation. But by no way it can represent the market sentiment as a whole. And IMO it doesn’t speak for the value of the model.