StrataNet vs LSTM vs GRU

Design Comparison

	LSTM	GRU	StrataNet
Domain	General	General	Market trading only
Hidden state	Opaque N-dim	Opaque N-dim	4-dim interpretable
State bounds	None	None	Hard bounds per dim
Gates	Input, forget, output	Update, reset	Update, reset + market projection
Output	Raw logits	Raw logits	LONG/SHORT/HOLD + regime + confidence
Parameters (default)	~12,000	~9,000	~5,000
Label source	Future price required	Future price required	Auto from STRATA teacher
Interpretable at inference	No	No	Yes — always

Hidden State Comparison

# LSTM hidden state — 32-dim opaque
h_lstm = [-0.12, 0.34, -0.07, 0.89, ...]   # What does dim-3 mean? Unknown.

# GRU hidden state — 32-dim opaque
h_gru  = [0.41, -0.22, 0.15, -0.63, ...]   # Same problem.

# StrataNet hidden state — 4-dim, always interpretable
h_strata = {
    "bias":        0.82,   # bullish conviction (0.82 > 0 = bullish)
    "momentum":    0.45,   # moderate breakout energy
    "trap_risk":   0.18,   # low trap risk — safe to act
    "uncertainty": 0.31,   # moderate uncertainty
}

Benchmark Results (Synthetic OHLCV)

Trained and tested on 5,000 synthetic OHLCV bars with realistic regime shifts. Labels generated by STRATA state machine teacher (same labels for all models = fair comparison).

Model	Test Accuracy	Sharpe	Max DD	Params
LSTM	~75%	varies	varies	~12,000
GRU	~74%	varies	varies	~9,000
StrataNet	~79%	varies	varies	~5,000

Full benchmark notebook: notebooks/benchmark_stratanet_vs_lstm_gru.ipynb

Why StrataNet Uses Fewer Parameters

StrataNet's 4-dim hidden state is an inductive bias — it forces the network to compress all market information into 4 meaningful dimensions rather than spreading across 32+ opaque units.

This is similar to how domain-specific architectures (e.g. ResNet for vision, Transformer for language) outperform generic MLPs by encoding domain knowledge into the architecture itself.

Teacher-Student Training

StrataNet uses knowledge distillation from the STRATA rule-based engine:

1. Rule-based STRATA processes each OHLCV window and generates action + regime labels
2. StrataNet learns to replicate these decisions as a differentiable neural network
3. Result: generalises the teacher's pattern recognition, faster at inference

This means no manual labeling — any OHLCV history you have is sufficient for training.