Yes, I can discuss the underlying technology. Much of it is based on building large vectors similar to word2vec but much more exhaustive in terms of rich, scored and ranked vectors. Comparing vectors for similarity is almost as important as how one constructs the vectors in vector space. Mimicking the way a human might manually construct a vector remains key.
Relying on vector similarity as opposed to direct keyword matches is an approach that's taken. Utilizing statistical and probabilistic approaches as opposed to "words and rules" remains important.
The corpus is currently just limited Reuters public company profiles & descriptions but I plan to include SEC filings as well.
Traders, investors, hedge funds can engage is quick information arbitrage with it. For example, a stock runs up 20% in minute, you insert a keyword or the symbol related to the stock that ran up, you then get other stocks (a targeted basket) that have sympathetic, symbiotic and parasitic relationships before any research analyst can uncover the connections - rising tide lifts all boats or a lowering tide lowers them. Ref: "Contagious Speculation and a Cure for Cancer: A Non-Event that Made Stock Prices Soar" -http://www0.gsb.columbia.edu/whoswho/getpub.cfm?pub=1555
The system can include tiered consumer/trader/investor subscriptions, licensing and net profit sharing with selected hedge funds. SeekingAlpha, StockTwits, Yahoo Finance etc as revenue partners. We're raising a bit of funding in the meantime while we also use it as a trading tool ourselves.
It's a prototype that's being moved into production this week or next.
Thanks for the additional info. Its always fun to hear a little of the back story when people are willing/able to share.
I have built both prototypes and production systems using a couple of different approaches to text analysis. Its always fun to hear a little of the back story when people are willing to share.
I wish you all the best in converting it into a solid business.
Extremely nice implementation and very good explanation of its use case. What would be interesting in the future once you're at the marketing stage would be to run through an example using historical data to provide me with an indication of my returns.
Thanks and I agree that historical performance metrics would be a great indication of returns.
What I'd like to do would be to create a system that processes 5 years worth of historical data, gathers all symbols that have had spikes up or down by 10-75%, have good volume and liquidity and are affected by either news or an earnings report. I could then auto-construct the baskets that surround each issue and calculate returns on the long side or short side.
I like this system because it executes on 4 critical levels:
1. Works in a good market or in a bad market due to short and long positioning that can be taken whereas most trading systems (and Technical Analysis) only really work in good markets.
2. You don't have to be late in terms of buying into (or shorting) the direct issue. You buy into or short the indirect issues, which enable protection and diversification.
3. Nobody really is doing this today.
4. It's based on information arbitrage and hidden connections, not numbers.
5 comments
[ 3.2 ms ] story [ 17.6 ms ] threadI am curious about your corpus source for making the connections; company filings, company websites, financial news, other - all of the above?
Are you also clustering topics rather than just doing keyword level connections?
If it is not a trade secret can you discuss what text mining approaches you take with this service?
How do you see investors using this?
What is your business model?
Relying on vector similarity as opposed to direct keyword matches is an approach that's taken. Utilizing statistical and probabilistic approaches as opposed to "words and rules" remains important.
See: http://cymetica.com/cymetica_about.html
and the following: http://genopharmix.com/biomimetic-cognition/in_silico_cognit...
The corpus is currently just limited Reuters public company profiles & descriptions but I plan to include SEC filings as well.
Traders, investors, hedge funds can engage is quick information arbitrage with it. For example, a stock runs up 20% in minute, you insert a keyword or the symbol related to the stock that ran up, you then get other stocks (a targeted basket) that have sympathetic, symbiotic and parasitic relationships before any research analyst can uncover the connections - rising tide lifts all boats or a lowering tide lowers them. Ref: "Contagious Speculation and a Cure for Cancer: A Non-Event that Made Stock Prices Soar" -http://www0.gsb.columbia.edu/whoswho/getpub.cfm?pub=1555
The system can include tiered consumer/trader/investor subscriptions, licensing and net profit sharing with selected hedge funds. SeekingAlpha, StockTwits, Yahoo Finance etc as revenue partners. We're raising a bit of funding in the meantime while we also use it as a trading tool ourselves.
It's a prototype that's being moved into production this week or next.
I have built both prototypes and production systems using a couple of different approaches to text analysis. Its always fun to hear a little of the back story when people are willing to share.
I wish you all the best in converting it into a solid business.
What I'd like to do would be to create a system that processes 5 years worth of historical data, gathers all symbols that have had spikes up or down by 10-75%, have good volume and liquidity and are affected by either news or an earnings report. I could then auto-construct the baskets that surround each issue and calculate returns on the long side or short side.
I like this system because it executes on 4 critical levels:
1. Works in a good market or in a bad market due to short and long positioning that can be taken whereas most trading systems (and Technical Analysis) only really work in good markets.
2. You don't have to be late in terms of buying into (or shorting) the direct issue. You buy into or short the indirect issues, which enable protection and diversification.
3. Nobody really is doing this today.
4. It's based on information arbitrage and hidden connections, not numbers.