How to Backtest a Stock Strategy Properly: Survivorship, Look-Ahead and Costs

A backtest applies a strategy's exact rules to past prices to see how it would have performed. Done properly, it uses data that includes companies that no longer exist, adjusts for splits and dividends, makes every decision with information that was available at the time, trades at realistic prices after costs, and compares the result with a fair benchmark. Done carelessly, it produces impressive numbers that vanish the moment you trade real money.

What a backtest can and cannot tell you

A good backtest can show that a rule is not worth trading, which saves money, and give a realistic picture of drawdowns, losing streaks and how often the rule trades. It cannot prove the future will look like the past. Treat a passed backtest as permission to paper trade or trade small, not as a promise.

Step 1: write the rules completely

If a rule needs your judgement, it cannot be tested. Write down:

• Universe: which stocks and how the list is chosen.
• Signal timing: decided on the daily close, traded at the next open.
• Entry, exit, stop and time limit.
• Sizing: fixed dollar amount, equal slots, or risk-based (see position sizing).
• Costs per trade.
• Pass/fail tests, written before you see a single result.

Step 2: use honest data

Survivorship bias

If your data only contains companies that exist today, it has quietly removed every company that went bankrupt, was delisted or shrank. Those are mostly losers, so any strategy that buys stocks looks better than it really was.

Two fixes: use point-in-time data that lists which companies were really in the universe on each date, delisted ones included; or, if you must use today's list, compare your rule with simply holding the same list, which carries the same bias. Only beating that basket shows that the rule adds something.

Corporate actions

• Splits: in a 4-for-1 split, a $400 share becomes four $100 shares. Unadjusted data shows a 75% crash that triggers every stop. Use split-adjusted prices.
• Dividends: price-only data understates buy-and-hold returns for dividend payers. Use total-return (dividend-adjusted) prices, or add dividends separately.
• Mergers, spin-offs, ticker changes and delistings: each needs handling, or positions vanish or jump in the data.

Step 3: avoid look-ahead bias

Look-ahead bias means using information that was not available when the decision was made.

• Price signals: if a signal uses today's close, you cannot also buy at today's close. Buy at the next open.
• Fundamentals: a company's fourth-quarter results are often published weeks after the quarter ends. Use each figure from its publication date, not the period end.
• Index membership: use the members of the index on each date, not today's members.
• Stops: if the day's low touches your stop, assume you are filled at the stop or, if the stock gapped through it, at the open.

Step 4: include realistic costs

Costs per trade include commission, the bid-ask spread, slippage and, where relevant, margin interest and borrow fees.

Taxes, too, are higher for frequent traders in many countries (see stock taxes and record keeping).

Step 5: judge results with the right measures

Measure	What it tells you
Number of trades	Whether the sample is big enough to mean anything
Average trade and profit factor (gross profits / gross losses)	Whether the edge survives costs
Win rate with average win and loss	The shape of returns and the losing streaks to expect
Compound annual return	Growth of the whole portfolio
Maximum drawdown	The worst peak-to-trough fall you would have lived through
Sharpe ratio	Return per unit of volatility, for comparing strategies with different risk
Results by year and by half of the period	Whether the edge is steady or came from one lucky stretch

Compare every result with at least two benchmarks: an S&P 500 index fund, and buy-and-hold of the same stock list.

Step 6: guard against overfitting

Overfitting means tuning rules until they fit the noise of the past. Warning signs and defences:

1. Too many settings. Every extra parameter makes it easier to fit noise. Prefer simple rules.
2. Fragile settings. If 47/193-day averages work but 45/190 and 50/200 do not, the result is probably luck.
3. Too many tries. Test 100 random rules and a few will look great by chance. Count and report every variation you tried.
4. No out-of-sample check. Split the data: develop on one period, then check once on another you have not looked at.
5. Rules from nowhere. Rules based on published research, written down before the first test, are less likely to be accidents of the data.

How DanioX tests stock strategies

The DanioX stocks backtest follows these principles and states its own limits. Its rules come from published research and were written down, untuned, before the first run. Signals are decided on a close and traded at the next open, stops fill when the day's low touches them, a fixed cost is charged on every buy and sell, and prices are adjusted for splits and dividends. Each strategy must pass checks fixed in advance, including enough trades and profits in both halves of the period. Because the universe is today's largest companies, the page measures every rule against an equal-weight basket of the same stocks as well as the S&P 500. The methodology page explains the approach across all DanioX markets.

Next: stock taxes and record keeping.