Strategy Card

Cost-Aware Selective Extrema Crypto Timing — strategy card

Spec ID: spec-cost-aware-selective-extrema-crypto-1784918823 · Generated: 2026-07-24 19:12 UTC

Cluster: Ml Timing · Sub Cluster: Cost-Aware Crypto Extrema Timing

One-line description

Long-only spot crypto timing model that enters selectively when ML forecasts indicate a probable local minimum, positive 4h net edge after costs, and low local-maximum risk. Positions exit on stop, target, timeout, adverse extrema risk, or expected-edge reversal.

Why this trade exists

Crypto markets can exhibit short-horizon mean reversion around local extrema, driven by liquidity shocks, forced flows, retail overreaction, and fragmented 24/7 market structure. The strategy attempts to monetize rebounds after statistically identifiable local lows while avoiding periods where the next local high or insufficient post-cost edge is likely.

The core hypothesis is not that every predicted trough is tradable, but that a calibrated, cost-aware abstention policy can filter for only the rare extrema events with enough expected 4h edge to survive spread/fee/slippage assumptions. Walk-forward retraining and calibration are used to reduce static-regime overfit, while long-only spot constraints avoid borrow and funding assumptions.

Algorithm

[code omitted from public view]

Parameters

Param	Value	Notes
Universe	Top 10 crypto by 1y dollar volume	Long-only spot universe
Data	1-minute OHLCV, resampled to 5min primary and 1h secondary	No synthetic execution prices; skip missing execution bars
Backtest window	2020-01-01 to 2025-10-08	Walk-forward monthly test folds
Initial training	18 months from 2020-01-01	Monthly retraining
Calibration	3 months	Thresholds selected on calibration fold only
Label horizon	4h	Forward return from next executable open to 4h-later open
Local extrema labels	40 bps min pullback/rebound	Centered extrema; radius 48 bars on 5min and 4 bars on 1h equivalent
Features	96-bar lookback	Return lags, volatility, range, volume z-scores/spikes, price location, distance to highs/lows, VWAP-proxy deviation, time features
Models	HistGradientBoostingClassifier + ExtraTreesRegressor	Balanced classifier weights; isotonic calibration; ExtraTrees with 300 trees
Entry thresholds	p_min grid 0.60-0.85; net-edge grid 21-100 bps	Selected by calibration net return after costs
Entry cost hurdle	31 bps primary round trip	Stress grid: 21, 31, 51 bps
Entry risk filter	maximum-risk probability < 0.35	Avoid predicted local highs/adverse regimes
Exit rules	-75 bps stop, +120 bps target, 48-bar timeout	Also exit on max-risk probability >= 0.55 or expected edge <= 0
Position sizing	Equal-weight capped	Max 10% NAV per symbol; max 3 concurrent positions
Leverage	Max 4.0 configured; no borrowing enabled	Entries skipped or resized to fit constraints
Execution	Next resampled bar open	Signal after completed bar; model exits also next bar open
Costs	31 bps primary round trip, half per side	Terminal liquidation cost included; separate broker commission fields also present

Look-ahead audit

#	Concern	Status
1	Centered local-extrema labels use future bars	Acceptable for supervised label construction only; labels are purged/embargoed and not used as live features
2	Feature normalization leakage	Mitigated: normalization fit on training fold only
3	Threshold overfit to test data	Mitigated: threshold selection uses calibration fold only, with test fold held out
4	Forward-return label overlap across splits	Mitigated: 4h purge horizon and embargo by bar after split
5	Execution timing	Mitigated: signals generated after completed resampled bar and executed only at next bar open
6	Missing or synthetic prices	Conservative: no synthetic execution prices; trades skipped or exited when required execution price is unavailable

Caveats / known limitations

• The reported backtest produced 0 trades and 0 symbols, so no performance inference can be drawn from the current run.
• The policy gate or data/universe availability likely filtered out all test-fold opportunities; this should be debugged before interpreting the strategy economics.
• Centered extrema labels are useful for training but can be fragile: small changes in radius or pullback/rebound threshold may materially alter labels.
• Short-horizon crypto execution is highly sensitive to fees, spread, market impact, exchange outages, and missing bars; the modeled 31 bps round-trip cost may still be optimistic or pessimistic depending on venue and size.
• ML classifiers may learn volatility/regime artifacts rather than persistent extrema structure, especially with rare-event labels and selective abstention.
• Long-only spot design avoids borrow risk but may underperform in sustained downtrends where rebound signals fail repeatedly or gates suppress activity.
• Stop/target evaluation from bar high/low is path-ambiguous; the implementation assumes adverse stop first on same-bar ties, but intrabar execution realism remains approximate.

Results

The current backtest summary is effectively a no-trade result: total return, Sharpe, Sortino, Calmar, volatility, drawdown, win rate, profit factor, VaR/CVaR, and average win/loss are all reported as 0 because the run generated 0 trades across 0 symbols. This should be treated as an implementation or gating outcome rather than evidence of alpha; the next research step is to verify universe loading, policy-gate pass rates, calibration trade counts, and threshold strictness.

Backtest metrics snapshot

Metric	Value
Total Trades	0
Symbols	0