KILLED · predicts the mean
Hypothesis
A recurrent LSTM trained on multivariate features (OHLCV, OI, funding) predicts the next-bar return well enough to trade directionally.
Math — what the loss actually optimises
Minimizing MSE on returns drives the output to the conditional mean:
$$ \min_\theta\ \mathbb{E}\big[(r_{t+1} - f_\theta(\mathbf{x}_t))^2\big]\ \Rightarrow\ f_\theta \to \mathbb{E}[r_{t+1}\mid \mathbf{x}_t]\approx 0 $$
Method
2-layer LSTM, 40 features, walk-forward train/validate, predict next 1h return; trade sign of prediction with costs.
Results
| Direction accuracy (out-of-sample) | 50.6% |
| Predicted |return| vs realized | shrinks to ~0 |
| Net of fees | − |
Because next-bar return is ~unpredictable, MSE training collapses the network toward predicting zero (the conditional mean). 50.6% directional accuracy ≈ coin flip, well under the fee hurdle. The architecture is irrelevant when the target has no learnable structure.
Deep nets do not manufacture signal. If the conditional mean of your target is ~0, an MSE-trained model will faithfully predict ~0. Pick a target with structure, not a fancier model.