Thread by @fchollet

There are lots of blog posts out there claiming to teach you how to use deep learning (typically RNNs) to predict stock prices, FX rates, BTC price, from past price data. Is it actually possible?

Well, mostly not. A thread.

All successful trading is information arbitraging: exploiting a piece of information others don't know (or won't/can't use). That's because market are dynamic: trading a signal erases the signal.

If any given source of information or model (e.g. Bollinger bands) becomes known to more than a few players, it stops being effective (unless you want to compete of how fast you can act on the signal).

For this reason, it is not possible to use publicly available price data and widely-known DL models to make money trading liquid & easy to trade assets.

Moreover, price data is a source of info that's already so squeezed that no amount of modeling improvement will help you.

There was likely a short period of time in the past where using LSTM represented a meaningful advantage. Likewise there was a point in time where a DL trading bot that looked at past BTC prices would have made money.

But not anymore, at least not at any meaningful scale.

That doesn't mean that price data doesn't contain information. Only that this information doesn't convert to successful strategies for an average player.

It's possible to show models that look like they would be successful, in theory. But you can't deploy them.

Remember, you need an advantage no one else has (usually a new dataset). It's purely a game of information arbitrage. You can't recycle the same information and models everyone has access to.

Unless you know you're special, don't even try. And if you have to ask, you're not.