Skip to content

STRATA — Market State Framework

A trainable, stateful neural network architecture purpose-built for market trading.

STRATA is a direct competitor to LSTM/GRU/Transformer for financial time-series, but with a fully interpretable hidden state and outputs designed specifically for trading decision support — not price prediction.

pip install strata-market

What is STRATA?

STRATA provides three levels of API — from foundation model to classic rule-based:

from strata import StrataFormer, StrataFormerConfig, StrataFormerPretrainer

# Step 1: self-supervised pretraining — no labels needed
cfg     = StrataFormerConfig(max_seq_len=256, n_layers=4, n_assets=10)
model   = StrataFormer(cfg)
trainer = StrataFormerPretrainer(model)
model   = trainer.pretrain(raw_candles_10000_bars, epochs=50)

# Step 2: fine-tune on specific asset
from strata import StrataFormerTrainer, StrataFormerDataset
dataset = StrataFormerDataset.from_candles(aapl_candles, seq_len=128, asset="AAPL")
trainer = StrataFormerTrainer(model=model, lr=1e-4)
model   = trainer.train(dataset, epochs=20)

# Step 3: variable-length inference
result  = model.predict_action(torch.randn(200, 5))   # any length!
# {"action": "LONG", "confidence": 0.71, "context_len": 200,
#  "state": {"bias": 0.82, "momentum": 0.45, "trap_risk": 0.18, ...}}

from strata import StrataNet, StrataNetTrainer, StrataNetDataset

# Train from OHLCV — no manual labeling needed
dataset = StrataNetDataset.from_candles(candles, seq_len=30, asset="AAPL")
trainer = StrataNetTrainer(asset="AAPL")
model   = trainer.train(dataset, epochs=30)
model.save("aapl_net.pt")

result = model.predict_action(x_tensor)
# {"action": "LONG", "confidence": 0.71, "regime": "TRENDING",
#  "state": {"bias": 0.82, "momentum": 0.45, "trap_risk": 0.18, ...}}

from strata import StrataModel, StrataTrainer

# Use pretrained (zero setup)
model  = StrataModel.from_pretrained("AAPL")
result = model.predict(candles)
# {"action": "LONG", "confidence": 0.71, "regime": "TRENDING", ...}

# Or train your own
trainer = StrataTrainer(asset="MSFT")
model   = trainer.train(windows, n_trials=100)
model.save("msft_model.json")

Key Differentiators

LSTM / GRU Transformer (GPT) StrataNet StrataFormer
Domain General General Trading only Trading only
Hidden state Opaque Opaque 4-dim semantic 4-dim semantic
Context window Fixed Variable Fixed 30 Variable 30–512+
Multi-asset ❌ ❌ ❌ ✅
Pretraining Supervised Next-token STRATA teacher Self-supervised (MBM)
Parameters Millions Billions ~5K ~15K–5M (scalable)