Changelog¶

Full changelog: CHANGELOG.md

v2.7.0¶

StrataFormer: trading foundation model — the backbone for trading-specific LLMs
Variable context window: 30 to 512+ bars (vs fixed 30 in StrataNet)
Multi-asset tokenizer: multiple tickers in one forward pass with learned asset embeddings
Causal self-attention stack: N × StrataCausalAttentionBlock (no future leakage)
Interpretable bottleneck: attention output → 4-dim [bias, momentum, trap_risk, uncertainty]
Masked Bar Modeling (MBM): self-supervised pretraining — no labels required
StrataFormerPretrainer: pretrain on raw OHLCV corpus
StrataFormerTrainer: supervised fine-tune with STRATA teacher labels
StrataFormerDataset: multi-asset dataset with 75% overlap windows
32 tests, all passing

Tick-level sense: sense() auto-detects bid/ask → spread-based anticipatory liquidity_above
sense_tick(): extended signals — spread_pressure + side_imbalance
GitHub Actions CI: automated tests on Python 3.10 + 3.11
Hugging Face Hub: pretrained weights at emylton/strata-net
MkDocs docs site: this site

StrataNet: PyTorch neural network architecture — direct competitor to LSTM/GRU/Transformer
4-dim interpretable hidden state: [bias, momentum, trap_risk, uncertainty]
StrataNetTrainer: gradient descent training with teacher-student knowledge distillation
StrataNetDataset: auto-label generation from STRATA state machine

StrataModel: JSON-based trainable model (train, save, load, predict, from_pretrained)
StrataTrainer: walk-forward coordinate optimizer
Pretrained weights for AAPL, TSLA, SPY, NVDA, QQQ