optiona-cnt21-nifty

CNT 2.1 — HF Inference Endpoint

Price-only credit spread profitability predictor for NIFTY options.

60 close-price candles → Chronos-2 base (768D) → MLP classifiers → bull/bear probabilities

Serves both spread widths per direction: width-100 (legacy keys, primary bull_probability/bear_probability) and width-200 (_w200 keys). One embedding pass feeds all classifiers.

Files

• handler.py — Endpoint handler
• requirements.txt — Python dependencies
• model_store/ — Classifier weights, scalers, metadata (built from the MLflow registry)
• model_store/deployment_manifest.json — registry versions, run IDs, hashes of this build
• .hfignore — Excluded from upload

No adapter directory needed — loads amazon/chronos-2 base model from HF hub.

Prepare & Upload

python -m deploy.build_model_store   # pulls @production versions from MLflow registry
cd hf_inference && python test_handler.py   # local smoke test

Then, from the project root, python upload_to_hf.py. A push is an immediate production release — the endpoint auto-tracks the latest revision. Verify locally first; after pushing run python test_hf_point.py and compare thresholds against deployment_manifest.json. (sync_assets.sh is the legacy path-based sync — deprecated, see DEPLOYMENT_AUDIT_CNT21.md for why.)

Input contract

POST JSON body:

{
  "inputs": [
    {"date": "2025-06-30 09:15:00", "close": 24150.5},
    {"date": "2025-06-30 09:18:00", "close": 24148.2}
  ],
  "parameters": {"context_length": 60}
}

Required columns

Minimum 60 rows (3-min candles), sorted ascending by date.

Parameters

Output contract

{
  "context_window_start": "2025-06-30 09:15:00",
  "context_window_end": "2025-06-30 12:12:00",
  "embedding_dim": 768,
  "bull_probability": 0.39,
  "bear_probability": 0.63,
  "bull_probability_w200": 0.41,
  "bear_probability_w200": 0.65,
  "predictions": {
    "bull_put_credit":       { "model_key": "bull_put_credit",       "target_column": "bull_put_credit_label",  "prob_label_1": 0.39, "threshold": 0.633, "y_pred_bin": 0, "y_pred_label": -1 },
    "bear_call_credit":      { "model_key": "bear_call_credit",      "target_column": "bear_call_credit_label", "prob_label_1": 0.63, "threshold": 0.627, "y_pred_bin": 1, "y_pred_label": 1 },
    "bull_put_credit_w200":  { "model_key": "bull_put_credit_w200",  "target_column": "bull_put_credit_label",  "prob_label_1": 0.41, "threshold": 0.656, "y_pred_bin": 0, "y_pred_label": -1 },
    "bear_call_credit_w200": { "model_key": "bear_call_credit_w200", "target_column": "bear_call_credit_label", "prob_label_1": 0.65, "threshold": 0.648, "y_pred_bin": 1, "y_pred_label": 1 }
  }
}

Current build thresholds: w100 bear_call 0.6268, bull_put 0.6326; w200 bear_call 0.6482, bull_put 0.6564 (see deployment_manifest.json).

Dependency note

• Keep torch aligned with HF base image to avoid ABI errors.
• No peft dependency needed (base model only, no LoRA adapter).