Research framework
Multi-Model Feature Engineering Framework
Regime · Volatility · Filter · Frequency · Memory · Decomposition
Six classical and modern quantitative models, each producing a complementary feature stream for the AlphaMind Prediction Engine. Deep learning does not replace the statistics here. It is stacked on top of them.
The premise
Before the first token
The prediction engine reads price and nothing else. That discipline holds because the price it reads has already been worked on. Before a single bar is tokenized, six quantitative models have passed over the same stream: the noise is filtered out, the series is decomposed, the current regime is inferred, the volatility ahead is estimated. What reaches the engine is not a raw chart but a chart with its context computed and attached.
The six are drawn from the quantitative literature rather than invented here. Each carries its citation, and each answers a question the foundation model should not have to relearn from scratch on every request.
The models
Six models, each feeding its own stream
Each one is published, cited and stated with the formula it runs. The last line of each entry is what its stream changes inside the engine.
01
HMM Regime
Regime · Hidden Markov Model
A probabilistic state-space model that infers latent market regimes (bull / bear / ranging) from observable price dynamics. Every downstream prediction is conditioned on the current regime — a forecast that ignores regime is, by definition, mis-specified.
P(sₜ | x₁:ₜ) ∝ P(xₜ | sₜ) · Σ P(sₜ | sₜ₋₁) · P(sₜ₋₁ | x₁:ₜ₋₁)
Rabiner (1989) · Hamilton (1989)
Provides the regime label that gates which prediction head the engine routes to.
02
HAR-RV
Volatility · Heterogeneous Autoregressive Realized Volatility
A long-memory volatility forecaster that aggregates realized variance at daily, weekly and monthly horizons. Volatility today depends on volatility across multiple time scales — a property GARCH alone misses.
RVₜ₊₁ = β₀ + β_d·RVₜ⁽ᵈ⁾ + β_w·RVₜ⁽ʷ⁾ + β_m·RVₜ⁽ᵐ⁾ + εₜ₊₁
Corsi (2009)
Calibrates the engine's confidence intervals — the prediction band widens or tightens with the vol forecast.
03
Kalman Filter
Filter · Recursive State-Space Estimator
An optimal recursive filter that fuses noisy observations with a dynamic model of the underlying state. It extracts the unobserved fair-price component from raw quotes — the engine sees a denoised signal, not market noise.
x̂ₜ|ₜ = x̂ₜ|ₜ₋₁ + Kₜ(yₜ − Hₜ x̂ₜ|ₜ₋₁)
Kalman (1960)
Pre-processes the OHLCV stream into a noise-reduced state vector before tokenization.
04
ATFNet
Frequency · Adaptive Time-Frequency Network
A neural architecture that learns features jointly in the time and frequency domains via a shared adaptive backbone. It captures periodic structure — intraday seasonality, session effects — that pure time-domain models systematically miss.
h = Adaptive[ Encoder_t(x) ⊕ Encoder_f(F{x}) ]
Liu et al. (2024)
Provides frequency-aware feature embeddings concatenated to the token stream.
05
Hurst Exponent
Memory · R/S Long-Memory Statistic
A scaling exponent estimated from the rescaled-range 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.
E[R(n)/S(n)] ∼ n^H, H = lim (log E[R/S]) / (log n)
Hurst (1951) · Mandelbrot & Wallis (1969)
Selects between trend-following and mean-reverting prediction heads inside the engine.
06
VMD
Decomposition · Variational Mode Decomposition
A non-recursive signal decomposition that splits a price series into band-limited oscillating modes around adaptively estimated central frequencies. Each mode is modelled separately, then re-aggregated.
min { Σₖ ‖∂ₜ[(δ + j/(πt)) ∗ uₖ(t)] e^{−jωₖt}‖² }
Dragomiretskiy & Zosso (2014)
Multi-band feature stream — high / mid / low frequency components predicted independently.
Division of labour
Six models split across three jobs
The six streams are not a committee voting on direction. Each one changes a specific thing about how the engine reads, and no two of them change the same thing.
Clean what the engine reads
Kalman Filter · VMD
Raw quotes carry noise the model should not spend capacity on. The Kalman filter extracts the underlying state from the stream, and VMD splits the series into frequency bands that are modelled separately. The engine reads a signal, not static.
Choose the prior it reads with
HMM Regime · Hurst Exponent
The same pattern means different things in different conditions. The regime model infers whether the market is trending, ranging or breaking down, and the Hurst exponent measures whether moves tend to persist or revert. Together they decide which of the engine's prediction heads a request is routed to.
Shape the band around the path
HAR-RV · ATFNet
HAR-RV forecasts volatility across daily, weekly and monthly horizons, and that forecast calibrates how wide the engine draws its band. ATFNet contributes the periodic structure that pure time-domain reading misses, from session effects to intraday seasonality.
Downstream
Every stream ends in the same place. What these six models produce is the feature set the AlphaMind Prediction Engine reads, and that model has a page of its own.
The Prediction Engine
