Regression and Classification, which is better in financial market price prediction?

I want to use a model to trade in finanical market.

which i have several features, like macd, rsi, or other common features.

and my target is to make a tradeable predict in every time point.

so my target can be:

1. yield in a fixed time laster, like, 30 min. yt = close(t+ws) - close(t)
2. futures price direction, which only can be 1(price up in the future) -1 (price down in the future)

these are difference between regression and classification.

Thanks

To sum it up: I think it comes down to the strategy and not the model itself.

In my opinion, classification is more appropriate for trading algos and here's why:

• In ordinary regression, what you are predicting is the exact value of the target variable (price change) from an infinite set of possibilites.
• In classification, however, you are predicting the class of the target variable from a finite list of possible classes. This can be simply predicting if the price is going up or down (two classes). Let's stick with two classes for the sake of simplicity in this post.

As you can see, predicting the exact value of a continuous variable with ordinary regression is much more challenging than predicting the correct class out of a limited number of classes. With two classes, you even have a 50% chance of being correct without even any modeling and just by pure chance. If you can improve the model to do slightly better than 50%, say, 55%, you can trade profitably over the long run.

From a practical perspective, what you most care about is getting the directional of price movement right (classification). You care less about if price is going up by 8% or 9% (regression)

The other advantage of classification is that you get a probability level associated with any prediction denoting confidence level in the prediction, which can be used for position sizing. So, for example, if I predict that tomorrow's price is going up/down with a probability of 70%, I may put, say, $$100 in that trade, while if the probability were 94%, I may put$$250 in the trade as I am more confident in my prediction.

Lastly, typically 50% of price changes are positive (up) and 50% negative (down). From the technical perspective, this works in favor of classification.

I see minimal value in just predicting the direction of price movement.

If you don't know the magnitude, you don't know if the profit resulting from making a trade would be wiped out by a trading fee or taxes.

If you don't consider the magnitude, you could do well nine times out of ten, only for an error on that tenth prediction to wipe out your profits.

I'm perfectly content to lose a little money nine times and then make a squillion dollars on my tenth attempt, but I am not willing to make money nine times only to lose it all because of one bad trade.