Write the hypothesis before the feature set
A useful hypothesis names a mechanism, universe, horizon, and expected response. Short-term reversal after a liquidity shock, cross-sectional momentum, and basis normalization are different claims. Inputs should connect to the claim and be available before the simulated decision.
Maintain an experiment ledger, including failed ideas and parameter changes. This reduces hindsight storytelling and makes multiple testing visible. A complex feature deserves stronger validation when its mechanism and timing are unclear.
Use point-in-time data and causality
Universe membership, listings, borrow access, liquidity, token supply, and venue availability should reflect what was known at each historical date. Every record benefits from both event time and available time. Bar labels, revised indexes, and forward-filled prices are common sources of accidental future information.
Normalize symbols, decimals, quote currencies, contract specifications, funding intervals, and timestamp policies. Data cleaning can improve quality without silently rewriting history. Keep the raw source, transformation version, and checks that excluded suspect intervals.
Separate forecasts from portfolio construction
An alpha score ranks expected relative return. The portfolio layer sets weights based on volatility, confidence, liquidity, factor exposure, turnover, and limits. This separation prevents favorable leverage choices from being attributed to predictive skill. Constraints should correspond to the claim of neutrality.
Model transaction cost as a function of asset, venue, size, and execution style. Penalize turnover and require marginal expected return to clear marginal cost plus uncertainty. Capacity analysis should show how edge changes as notional grows.
Validate across time and alternatives
Use chronological splits, walk-forward testing, and a final holdout. Purge overlapping labels where necessary. Compare sophisticated models with simple baselines using identical costs and constraints. Seek broad parameter regions rather than a single optimum.
Stress spreads, impact, funding, latency, missing data, delistings, and factor shocks. Report gross and net performance, turnover, tail losses, drawdowns, exposure, and error bars. A strategy that depends on one asset or one month is evidence for further investigation, not deployment.
Monitor model and data drift
Production monitoring compares feature distributions, forecast calibration, realized cost, exposure, and residual returns with research expectations. A model can reduce risk, return to shadow mode, or retire when evidence changes. Automatic retraining should not automatically authorize new behavior.
Version datasets, features, models, parameters, and approvals. Every target portfolio should be reproducible from an input snapshot. This connects model research to the order and reconciliation controls required by a live system.