What it is

Backtesting is the process of applying a fully specified set of trading rules to historical price data to estimate how the strategy would have performed in the past. You feed the rules - entry condition, exit condition, position sizing - through past data bar by bar, record every simulated trade, and compute statistics such as win rate, average win, average loss, expectancy, and drawdown.

It is important to be precise about what a backtest actually shows. A backtest shows how a specific, mechanical rule set would have behaved on one particular history. It does not show how you will do in the future, it does not prove the strategy has an edge, and it certainly does not show how a discretionary trader who improvises will perform. A backtest is a hypothesis-testing tool, not a crystal ball. Its honest output is a probability statement: if the future resembles the tested past and you execute the rules exactly, here is the kind of behaviour you might expect.

How it works

A disciplined backtest has a few non-negotiable parts.

• Fully mechanical rules. Every decision must be reducible to data the test can evaluate. "Buy when it looks strong" cannot be backtested; "buy when the close crosses above the 50-day moving average" can.
• Realistic costs. Commissions, the bid-ask spread, and slippage must be subtracted from every trade. Ignoring costs is the fastest way to turn a losing system into a winning-looking one.
• A large, representative sample. Enough trades, across enough different market conditions, that the results are not an accident of one lucky stretch.
• Out-of-sample data. A portion of history reserved and never looked at during design, used once at the end to check whether the rules generalise.

The workflow is: design the rules on one slice of history (the in-sample period), then run them unchanged on a separate slice the rules have never seen (the out-of-sample period). If performance holds up out-of-sample, you have weak but real evidence of an edge. If it collapses, the in-sample result was probably an illusion.

How to read it

Reading a backtest well means reading it skeptically, because two specific biases routinely produce results that are far too good to be true.

The first is overfitting (also called curve-fitting). Overfitting is tuning a strategy so tightly to the historical data that it captures the noise - the random, non-repeating quirks - rather than a genuine, repeatable pattern. The classic symptom is a strategy with many parameters, each optimised to a precise value ("buy on the 14-day RSI below 27, sell above 73, but only on Tuesdays"), that posts a spectacular historical equity curve and then fails immediately on new data. The more knobs you turn and the more you re-optimise after seeing the results, the more you are fitting history rather than discovering a rule. Defences: prefer few parameters, favour round and robust values over precisely tuned ones, and demand that the strategy survive on out-of-sample data it was never optimised against.

The second is look-ahead bias, the most insidious flaw because it can be invisible. Look-ahead bias occurs when the backtest uses information that would not have been available at the moment of the simulated decision. Examples that quietly inflate results:

• Using a day's closing price to decide a trade that the rules say is taken during that same day.
• Filtering the universe to companies that still exist today, silently excluding firms that went bankrupt - a related flaw called survivorship bias.
• Incorporating financial figures (earnings, revisions) at the date they refer to rather than the later date they were actually published.

Look-ahead bias makes a strategy appear to predict the future when it is really just peeking at it. Because the bias is in the code or the data handling rather than the visible results, the only reliable defence is to enforce that every decision at bar t uses only data available at the close of bar t or earlier.

A worked illustration: suppose a moving-average crossover system shows a 90% annual return in backtest. Before believing it, ask: were costs and slippage deducted? Were entries taken on the next bar's open, or suspiciously at the signal bar's close? How many parameters were tuned, and was anything left out-of-sample? Was the universe frozen to today's surviving stocks? More often than not, one of these questions explains most of the apparent magic.

Two further habits separate honest backtesting from self-deception. The first is parameter robustness testing: instead of reporting only the single best setting, vary each parameter slightly and look at the surrounding results. A genuine edge sits on a broad plateau - the 48-, 50-, and 52-day moving average all work roughly as well. An overfit result sits on a lonely spike - only the 50-day works and 49 or 51 collapse - which is the signature of having fit noise. The second is walk-forward validation, a rolling version of out-of-sample testing: you optimise on an early window, test on the next, then slide both windows forward and repeat. If the strategy keeps performing on each freshly unseen window, the edge is far more credible than a single in-sample fit, because the test continually confronts the rules with data they were never tuned on.

It also helps to think about how many things you tried. If you tested two hundred variations and report the best one, that best result is partly luck - with enough attempts, some configuration will look great on any history by chance alone. This is why a strategy should ideally be specified from a clear economic or behavioural rationale before the search, not reverse-engineered from whatever curve looked prettiest afterwards.

Strengths & limits

The strength of backtesting is that it forces precision and provides a sample. To backtest at all, you must define your edge mechanically, which alone eliminates a great deal of vague thinking. A clean backtest on out-of-sample data, with realistic costs and few parameters, is genuine evidence - far better than an opinion or a single memorable trade.

The limits are severe and must be respected. A backtest can only test what already happened; a regime the data never contained (a novel crash, a new volatility regime) is invisible to it. Even a clean test is a sample, so it carries statistical uncertainty. And the entire exercise is fragile to the biases above: overfitting and look-ahead bias do not merely add noise, they systematically push results in the flattering direction, which is exactly the direction that loses real money. Treat every impressive backtest as guilty until proven innocent, validate forward, and size positions as if the true edge is smaller than the test suggests.

There is one more discipline that protects you in live trading: forward testing, also called paper trading or a forward log. After a backtest passes, you run the rules in real time on new, unfolding data - without risking capital, or with a deliberately tiny size - and compare the live results to the backtest's expectations. Because forward testing happens on data that did not exist when the rules were written, it is structurally immune to both look-ahead bias and overfitting. A strategy that backtests beautifully but quietly degrades in forward testing is telling you the historical result was at least partly an artefact. Honest evaluation therefore moves in a sequence - in-sample design, out-of-sample check, walk-forward validation, then forward testing - with each stage exposing the rules to data they have never seen, and with full position size earned only after the strategy has survived all of them.

Key takeaway: A backtest shows how a fully mechanical rule set would have behaved on one history, not how you will do; its results are only trustworthy after you have deducted realistic costs, limited overfitting, and rigorously excluded look-ahead bias.