futures-foundation-model

Futures Foundation Model (FFM) — Pretrained Backbone v2

A pretrained transformer backbone for futures market structure and regime classification. Trained on ~2.3M bars across 6 instruments (Oct 2020 – Apr 2026).

Model Files

Links

• GitHub: johnamcruz/Futures-Foundation-Model
• Fine-tuning framework docs: README

Usage

pip install git+https://github.com/johnamcruz/Futures-Foundation-Model.git

from huggingface_hub import hf_hub_download
from futures_foundation import FFMConfig, FFMBackbone

# Download weights
path = hf_hub_download(repo_id="johnamcruz/futures-foundation-model", filename="best_pretrained.pt")

config = FFMConfig()
backbone = FFMBackbone(config)
backbone.load_pretrained(path)

embeddings = backbone(features_tensor)  # (batch, 256)

Fine-Tuning a Strategy

from futures_foundation.finetune import StrategyLabeler, TrainingConfig, run_walk_forward

class MyStrategyLabeler(StrategyLabeler):
    @property
    def name(self): return 'my_strategy'

    @property
    def feature_cols(self): return ['zone_height', 'entry_depth', 'risk_norm']

    def run(self, df_raw, ffm_df, ticker):
        features_df, labels_df = my_signal_logic(df_raw, ffm_df)
        return features_df, labels_df

See the full documentation for the complete fine-tuning framework.

Pretraining Details

• Architecture: 6-layer transformer encoder, 8 heads, 256-dim, 512 FFN
• Input: 68 ATR-normalized continuous features + candle_type embedding (69 total)
• Sequence length: 96 bars (8 hours of 5-min data)
• Instruments: ES, NQ, RTY, YM, GC, SI
• Training data: ~2.3M bars, Oct 2020 – Apr 2026
• Pretraining tasks: Regime (4-class), Volatility State (4-class), Market Structure (2-class), Range Position (5-class)

Context Heads (best_pretrained.pt)

The pretrained checkpoint includes 4 frozen context heads that expose explicit market state at inference:

These 15 dimensions are concatenated into the fusion layer during fine-tuning, giving the signal head named handles on market state without relying on implicit encoding.

Why Use the Pretrained Backbone

• Regime changes don't require retraining — context heads adapt at inference as market conditions shift
• Adding new data is just a re-run — backbone representations are stable; only re-run strategy fine-tuning when new bars arrive
• One backbone, unlimited strategies — CISD, SuperTrend, ORB, breaker blocks all fine-tune on the same backbone
• 5+ year training window covers COVID crash, 2021 melt-up, 2022 bear, 2023 recovery, 2025 volatility — new regimes map to existing embedding space

License

Apache 2.0