From Hypothesis to Testable Rules

A trading strategy begins with a falsifiable market hypothesis, not an indicator combination. A useful hypothesis explains why a repeatable behavior might exist. For example: after a period of unusually low volatility, a high-volume breakout aligned with the long-term trend tends to continue because resting orders are triggered and under-positioned traders enter late. This claim can be tested. By contrast, “buy strong coins when the chart looks bullish” cannot be tested consistently because strong and bullish are undefined.

Translate the hypothesis into a decision that two independent traders would execute identically from the same candles. Specify the instrument universe, timeframe, session convention, setup, entry, initial stop, profit exit, time exit, and invalidation condition. Also define whether decisions use candle closes or intrabar prices. Every calculation must use information available at the decision timestamp; otherwise, the test contains look-ahead bias.

Consider a BTC-USDT four-hour momentum strategy using UTC candles. A valid setup requires the close to be above the 200-period exponential moving average, ATR over 14 completed candles divided by price to be no more than 1.8%, and the current close to exceed the highest high of the previous 20 completed candles. Breakout volume must be at least 1.3 times the median volume of the previous 20 candles. These thresholds are examples to validate, not universal constants.

Entry, Exit, and Invalidation Design

The entry occurs at the next four-hour candle open after a qualifying signal. This avoids the unrealistic assumption that an order can be placed at the same closing price that confirms the setup. The initial stop is 1.5 ATR below the actual entry. One unit of risk, or 1R, equals entry price minus stop price for a long trade. The profit target is 2R above entry, and any position still open after 12 candles is closed at the next available price.

A stop and a strategy invalidation are related but different. The stop controls loss on one trade. Strategy invalidation determines when evidence no longer supports deploying the system. An individual stop-out does not invalidate a positive-expectancy strategy. A reasonable operational rule might pause the strategy if live results produce a 12R drawdown, if average execution cost exceeds the backtest assumption by 50%, or if 30 consecutive eligible signals show expectancy below minus 0.10R. These thresholds must be selected before observing the live outcome.

Position size should make losses comparable across volatility regimes. With account equity of 20,000 USDT and risk per trade of 0.5%, the risk budget is 20,000 × 0.005 = 100 USDT. If entry is 60,000 and the stop is 58,800, raw distance is 1,200 USDT per BTC. Ignoring costs, size is 100 divided by 1,200, or 0.0833 BTC. If estimated round-trip fees and slippage equal 0.16% of entry price, execution cost is approximately 96 USDT per BTC. Cost-adjusted size is 100 divided by 1,296, or 0.0772 BTC. Leverage changes required margin, not the defined loss distance.

Building a Credible Backtest

Use clean historical candles from venues that match the intended execution environment. Crypto prices, volumes, fees, funding rates, and liquidity differ across exchanges. Record the symbol, exchange, contract type, timeframe, date range, fee tier, order type, assumed slippage, and funding treatment. For perpetual futures, excluding funding can materially overstate returns when positions remain open across funding timestamps.

Process signals chronologically and prevent overlapping positions unless the live strategy will allow them. Apply trading fees on entry and exit, adverse slippage, and realistic fills. If both stop and target fall inside the same candle, candle data alone cannot reveal which occurred first. Resolve this using lower-timeframe data or a conservative rule that records the stop first. Silently selecting the target introduces favorable bias.

Divide the data chronologically rather than randomly. One practical structure is 60% development, 20% validation, and 20% untouched final testing. Parameters are designed on the first segment, compared on the second, and assessed once on the final segment. A stronger approach uses walk-forward testing: optimize or select rules on a fixed historical window, test on the next period, then roll both windows forward. This exposes the strategy to changing volatility and trend regimes.

Parameter robustness matters more than the single best result. If a 1.50 ATR stop is profitable but 1.40 and 1.60 are deeply negative, the apparent edge may be noise. Test a neighborhood of plausible values and prefer a stable plateau. Avoid testing hundreds of variations and reporting only the winner. Each additional variation increases the probability of discovering a historical coincidence.

Reading Performance Quantitatively

Start with trade-level returns expressed in R because R normalizes trades by their initial risk. Suppose a test contains 120 trades: 54 wins and 66 losses. The win rate is 54 divided by 120, or 45%. If average winner is 1.65R and average loser is minus 1.00R, gross expectancy is 0.45 × 1.65 minus 0.55 × 1.00 = 0.1925R per trade. A strategy can therefore be profitable despite losing more often than it wins.

Profit factor compares gross wins with gross losses. Gross wins equal 54 × 1.65R = 89.1R, while gross losses equal 66 × 1.00R = 66R. Profit factor is 89.1 divided by 66, or 1.35. If average fees, slippage, and funding total 0.08R per trade, net expectancy becomes 0.1925R minus 0.08R = 0.1125R. At 0.5% account risk per trade, the approximate arithmetic expectation is 0.1125 × 0.5% = 0.05625% of equity per trade before compounding.

Do not judge the system by net profit alone. Record maximum drawdown, longest losing sequence, exposure, average holding time, trade frequency, and performance by market regime. A 30% return with a 28% drawdown has different career implications from a 22% return with an 8% drawdown. Sharpe and Sortino ratios can help, but their values depend on sampling frequency and assumptions about return distributions. For an event-driven system, the equity curve, drawdown duration, and R distribution often communicate risk more directly.

Forward Testing and Acceptance Criteria

Backtesting establishes historical plausibility; forward testing checks whether rules, data, and execution survive real-time conditions. Run the complete process on unseen incoming signals using paper trading or minimal capital. Orders must be recorded when the signal occurs, not reconstructed later. Capture intended price, actual fill, spread, slippage, fees, funding, timestamp, position size, stop changes, exit reason, and any missed or rejected order.

Choose acceptance criteria before forward testing. For example, continue until at least 40 eligible signals or 12 weeks have passed, whichever is longer. The strategy proceeds to limited deployment only if net expectancy remains above zero, realized profit factor is at least 1.15, maximum drawdown stays below 10R, and median slippage does not exceed the modeled value by more than 25%. These are governance thresholds, not proof that future profits are guaranteed.

Compare forward results with a range, not an exact backtest average. With only 40 trades, sampling uncertainty is large. Inspect whether win rate, average win, average loss, costs, and signal frequency are directionally consistent. A lower win rate may be harmless if winners are larger; a similar gross expectancy may still be unacceptable if execution costs doubled. Investigate deviations before changing rules.

Avoiding False Confidence

The main threats are look-ahead bias, survivorship bias, overfitting, incomplete costs, and discretionary overrides. Survivorship bias appears when today’s liquid token list is tested in earlier years, excluding assets that failed or were delisted. Overfitting appears when rules are repeatedly adjusted to repair every losing period. Discretionary overrides make the tested strategy different from the strategy actually traded.

Maintain a versioned strategy specification and change only one justified component at a time. A revised rule creates a new strategy version that requires fresh validation; it does not rewrite the old test. Keep the final test segment untouched until the complete rule set is frozen. If it is repeatedly consulted during development, it becomes part of the training data and a new out-of-sample period is required.

A validated strategy is not one with the highest historical return. It is one with an explicit economic hypothesis, reproducible rules, realistic execution assumptions, positive net expectancy, tolerable drawdowns, robustness across nearby parameters and regimes, and forward behavior consistent with historical uncertainty. Sustainable trading comes from controlling the testing process and risk exposure, not from eliminating losing trades.