54% Accuracy and Still Losing Money? You’re Optimizing the Wrong Objective

Shenggang Li13 min read·Just now

A first-principles explanation of the prediction–decision gap (Bandits → MDP RL), with a practical stock-trading example and AI prompt templates

I once saw a trading model that looked impressive on paper.

Gradient boosting (XGBoost).
A clean feature set: momentum, volatility, volume, macro proxies.
A backtest that reported 53–55% directional accuracy.

Then the real question hit:

If it predicts up more often than down… why doesn’t it make money?

The team’s answer was predictable:

Markets are random.
Costs are high.
It needs more features.

All true.

None was the real lesson.

The real lesson is a first-principles one:

Trading is not a prediction problem.
Trading is a decision problem.

Supervised ML is built to imitate labels.

But markets don’t give you “the correct action” label.

Markets give you something harsher:

You act. Then reality grades your action.

This article was originally published on Trading Tag and is republished here under RSS syndication for informational purposes. All rights and intellectual property remain with the original author. If you are the author and wish to have this article removed, please contact us at [email protected].