Holdout helper (train / holdout)

Practical anti-overfit around the research export lane — no ML.

Idea

1. Split bars into contiguous train then holdout
2. Run export_sma_sweep on train
3. Re-score the train top-K (fast, slow) pairs on holdout only
4. Report gap_best, mean_gap_top_k, overfit_risk, promote_ok

Schema: mn.holdout_report.v1.

API

from monte_neo.backtest import ExecutionModel, holdout_sma_sweep, synthetic_ohlcv

ohlc = synthetic_ohlcv(50_000, seed=42)
model = ExecutionModel(commission_bps=5.0, slippage_bps=5.0, warmup_bars=50)
report = holdout_sma_sweep(
    ohlc["open"], ohlc["high"], ohlc["low"], ohlc["close"],
    combos=64, top_k=5, model=model, device="auto",
)
print(report["metrics"])

Enrich HeuristicPolicy:

from monte_neo.policy import build_research_state, HeuristicPolicy

state = build_research_state(train_export_dict, holdout_report=report)
decision = HeuristicPolicy().decide(state)
# high holdout gap → promote blocked, lean run_mc / reject

CLI

monte-neo --holdout-sma --holdout-bars 20000 --holdout-combos 32

Synthetic smoke only — for real bars call the Python API.

Promote modes

Mode	Promote when
holdout_positive (default)	train_best > 0 and holdout_at_best > 0
strict	same, and overfit_risk != "high"

Large gaps still set overfit_risk / worth_mc_stress but no longer block the default mode.

Label log (for a future LocalScorer)

monte-neo --holdout-sma --holdout-log ~/mn_labels.jsonl --human-label accept

Or in Python:

from monte_neo.policy import append_research_label
append_research_label("labels.jsonl", holdout_report=report, decision=decision, human_label="reject")