Time Series

Sequence Models for Forecasting

A controlled benchmark of Transformer, state-space, and temporal-convolution models against classical time-series baselines for short-horizon market forecasting, with matched training and ablations.

2026In progress

Pits Transformer encoders, S4 or Mamba state-space models, and temporal convolutional networks against AR, Ornstein-Uhlenbeck, VECM, and GARCH baselines on returns and order-flow signals.

Evaluates out of sample by information coefficient, R-squared, directional accuracy, and the Sharpe of a simple signal-to-position rule, with ablations over context length, depth, and training horizon.

Runs first on synthetic stochastic-volatility paths where the truth is known, then on real market data, with fixed seeds and logged configs for every run.