4 ms·
I was simply using it to predict relatively short-term (between a few hours to 2-3 weeks) movements in the stock. I wasn't using it to even predict the magnitud
by crazypyro 12y ago
I was simply using it to predict relatively short-term (between a few hours to 2-3 weeks) movements in the stock. I wasn't using it to even predict the magnitude of the rise/fall beyond a few "tiers". Think like small/medium/large magnitudes and that's not far off.
With that said, I had between 60-75% success rate on predicting the direction, depending on how I set different things like requiring multiple sources, confidence required to execute prediction, NLP tweaking, etc. The tiers was worse, but I think it was around 40-45% on the last iteration. Obviously, I would want to get better at predicting magnitude before I used it in any real life trading. It also was highly dependent on the company/industry. It performed better on something like Apple or Google, but worse on some smaller companies in industries that are not near as news heavy.
The main goal of the project was just to satisfy my curiosity with algo trading, web scraping and data analysis while giving me another cool project to talk about in interviews, as I'm a college student. I used scrapy, NLTK, pandas mostly.
- nether 12y agoI think a 60% success rate is phenomenal, given that card counting only nets a 0.5-1.0% edge over the house. You've got yourself a 20% edge.
- notahacker 12y agoDon't forget that as the OP pointed out, they tested on direction rather than magnitude; if the losses from the predictions where it got the direction of movement wrong are big enough it might be no edge at all, or a negative ROI.
- lordnacho 12y agoIf it's really 60% and it works over the general universe of shares, you've built a money machine.
- omegaham 12y agoRemember that the general direction of the market right now is upward. This means that these day-trading algorithms need to beat the typical rate of return for, say, a diversified index fund. Many of them don't.
- lordnacho 12y agoWell I'm assuming the guy has thought about that, and his dataset is long enough to be meaningful. He didn't say how much the return varied, which you need for the comparison with the index to make sense.
- kasey_junk 12y agoNope because it doesn't account for risk.
- lordnacho 12y agoThe fact that it works over the universe accounts for the risk. If you have a 60% chance of being right, doing it over a few thousand stocks will make your daily return extremely steady, thanks to the law of large numbers. Of course there are correlations and such things to think about, but it's all controllable with that kind of edge.
- kasey_junk 12y agoNo it still doesn't. Because you have to allocate resources to the positional bet and one side of that bet has virtually limitless risk attached to it and the other has limited upside.
- lordnacho 12y agoWhat side has limitless risk? He didn't say what he was up to.
- kasey_junk 12y agoAll trades have 2 sides. Getting in and getting out. If your predictions include "I have a high degree of confidence that this instrument is about to go down" the way you enter that trade is by shorting it. Shorting has no true floor to the amount of money you can lose. This isn't true of going long, as an instrument can only go to zero. There is a natural imbalance between the max risk associated with shorting vs going long (and there are lots of other risks associated with trading, this is very simplistic) that means that no one who has done this sort of thing before uses a predictive ability without risk as a serious basis for a trade. In fact, predictive ability is not that special. There are trivial algorithms that can predict movements with high levels of accuracy (much higher than 60% for instance), but don't perform well on a risk adjusted basis in modelling or in reality.