Regime · Volatility · Filter · Frequency · Memory · Decomposition
Six classical and modern quantitative models, each producing a complementary feature stream that feeds the AlphaMind Prediction Engine. We don't replace statistics with deep learning — we stack them.
Markets are heterogeneous — regime, volatility, microstructure noise, and frequency all live on different time scales. One model can't see all of them well.
Six independent models with non-overlapping inductive biases yield strictly less correlated errors. The engine sees a richer, decorrelated feature space.
Each module's output is mathematically traceable. When the engine's forecast moves, we can attribute it back to which feature stream changed.
Each model below produces a feature stream that gets fed into the AlphaMind Prediction Engine — together they form the engine's input embedding.
A probabilistic state-space model that infers latent market regimes (bull / bear / ranging) from observable price dynamics. We condition every downstream prediction on the current regime — a forecast that ignores regime is, by definition, mis-specified.
Provides the regime label that gates which prediction head the engine routes to.
A long-memory volatility forecaster that aggregates realized variance at daily, weekly, and monthly horizons. It captures the fact that volatility today depends on volatility across multiple time scales — a property that GARCH alone misses.
Calibrates the engine's confidence intervals — the prediction band widens or tightens with HAR-RV's vol forecast.
An optimal recursive filter that fuses noisy observations with a dynamic model of the underlying state. We use it to extract the unobserved 'fair-price' component from raw quotes — feeding the engine a denoised signal instead of raw market noise.
Pre-processes the OHLCV stream into a noise-reduced state vector before tokenization.
A neural architecture that learns features jointly in the time and frequency domains via a shared adaptive backbone. It captures periodic structures (intraday seasonality, session effects) that pure time-domain models systematically miss.
Provides frequency-aware feature embeddings that get concatenated to the token stream.
A scaling exponent estimated from the rescaled-range (R/S) statistic of a price series. H = 0.5 ⇒ random walk; H > 0.5 ⇒ persistent / trending; H < 0.5 ⇒ mean-reverting. It tells the engine which prediction prior to use.
Selects between trend-following and mean-reverting prediction heads inside the engine.
A non-recursive signal decomposition that splits a price series into a small set of band-limited oscillating modes around adaptively estimated central frequencies. We model each mode separately, then re-aggregate — a decomposition-then-forecasting pipeline.
Multi-band feature stream — high / mid / low frequency components are predicted independently.
No complex jargon. No noise. No confusing indicators
Just clear, actionable signals.