Strategy research & validation

“Buy strong breakouts” is an observation, not a testable strategy. A falsifiable rule names the universe, timestamp, inputs, entry, size, exit, invalidation, and conditions under which no trade occurs. Every input must have been available at the decision time.

Write the failure condition before seeing results. For example: after costs, the lower-confidence estimate of expectancy must remain above zero on untouched data; otherwise the hypothesis fails. This prevents a weak idea from surviving through endless reinterpretation.

A backtest must reconstruct the information set available at each decision. Using today's surviving coins deletes failures; using a candle's close to trade at that same close can leak the completed bar; using revised macro or on-chain data substitutes a later history for the original release.

Define eligibility through time, retain delistings, align event and availability timestamps, document missing data, and distinguish exchange time from local display time. Keep raw inputs immutable and build cleaned research tables through reproducible transformations.

Use training data to develop the rule and validation data to make limited model choices. Freeze the complete specification before opening the final test. If you inspect the test and tune again, it has become training data—record the attempt and obtain a genuinely new forward sample.

Time-series splits must preserve order. Randomly shuffling observations can let nearby regimes and overlapping labels leak across folds. Walk-forward testing repeatedly estimates on the past and evaluates on the next unseen period, exposing regime dependence and operational drift.

Backtests often mark entries and exits at a candle midpoint or close. A real strategy crosses or joins a spread, waits in a queue, receives partial fills, moves the book, pays venue fees, and may pay funding, gas, borrow, or failed-transaction costs.

Use a conservative executable-price model tied to side, order type, size, and contemporaneous liquidity. Stress the estimate above its fitted value. A strategy that survives only at zero cost has not demonstrated a tradable edge.

A lone parameter combination with spectacular results surrounded by losses is usually fragile. Prefer a broad plateau where nearby parameters retain similar behavior. The goal is not the highest historical Sharpe ratio; it is evidence that the rule captures a repeatable mechanism rather than sample noise.

Stress symbols, venues, time periods, start dates, costs, latency, missing trades, and small rule changes. Explain why the edge should exist, who pays it, why competition has not removed it, and what capacity or behavioral constraint keeps it alive.

Forward validation records signals as time passes without rewriting the scorecard. Minimal live deployment then measures what simulation cannot: actual fills, queue behavior, latency, rejects, funding, operational errors, and your own adherence.

Before launch, define expected ranges and alerts for signal count, win/loss distribution, slippage, costs, drawdown, exposure, and rule violations. Pause when assumptions break; investigate before changing the rule; retire an edge when new evidence no longer supports it.