TFT Loved Less Training Data

The setup: NeuralForecast Tier-1 baselines (NBEATSx, NHITS, PatchTST, TFT) trained on commodity returns with min_train=2500 walk-forward folds — about ten years of training data per fold. I wanted to test whether more recent training history mattered more than longer history, so I dropped min_train to 1500 (~6 years), giving up training data to extend the OOS window.

The expectation was modest movement: maybe ±0.1 Sharpe across architectures. What landed:

Mean Δ across all 16 cells: +0.36. Mean Δ for TFT specifically: +1.07.

The other architectures barely moved. NBEATSx and PatchTST occasionally got worse. NHITS was random. Only TFT — the largest, most parameter-heavy of the four — got dramatically better with less training data.

This is the opposite of the “big models need more data” intuition. The likely mechanism: TFT’s variable-selection-network and attention layers are sensitive to fold composition. Training on the most recent 6 years lets it learn from a market regime closer to the OOS test period (post-2018: low-yield, high-vol, COVID, inflation cycle). Training on the full 10 years pulls in the 2010-2017 regime, which dilutes the signal TFT actually needs to capture.

Smaller architectures (NBEATSx, NHITS) don’t have enough capacity to chase regime-specific signal — they fit the long-run features, and longer training mostly helps them by reducing variance. TFT has the capacity, so giving it more relevant data outweighs giving it more total data.

The headline impact on the ensemble: portfolio_v11 (which now includes the extended-OOS TFT predictions) lifts honest walk-forward Sharpe from 1.94 → 1.98 with these candidates added. Most of the lift comes from the new extoos_TFT slots in the v7 candidate pool.

Lesson I’m taking: don’t default to “more training data is better.” For big-architecture small-dataset cases — which is most quant ML — recency matters more than total volume. The right min_train is a hyperparameter worth sweeping per architecture.