Neural network architectures for financial time series
What architectures are people finding useful?
My experiments:
- LSTM/GRU: Decent for univariate series, overfit on multivariate
- Temporal Fusion Transformer: Best interpretability, strong performance
- TCN (Temporal Convolutional Networks): Fast training, competitive accuracy
- N-BEATS: Great for pure forecasting tasks
35 Replies
I disagree about the GARCH approach. In my experience, realized volatility estimators (like the Rogers-Satchell estimator) outperform parametric models.
Be careful with the Kelly criterion for position sizing. Full Kelly is way too aggressive. I use quarter-Kelly in practice.
Have you considered using PCA to reduce the dimensionality of the feature space? I found that the first 10 components capture 80%+ of the variance.
The competition scoring docs could definitely be clearer. I spent 2 hours debugging what turned out to be a normalization issue.
One more thing: the scoring engine uses a held-out test period that you never see. So your validation score is the best you can do.
Good point about overfitting. My rule of thumb: never trust a backtest with fewer than 500 observations in the out-of-sample period.
Be careful with the Kelly criterion for position sizing. Full Kelly is way too aggressive. I use quarter-Kelly in practice.