Backtesting

Context & Mechanics

Definition and process

Backtesting answers the question: “What would this strategy have earned in the past – and what would it have cost?” The process has four steps: first, define the strategy fully rule-based (entry, exit, position size – no discretion). Second, obtain clean historical data covering the target market and period. Third, simulate the rules trade by trade – including fees, bid-ask spread and realistic slippage. Fourth, condense the results: expectancy, profit factor, win rate, max drawdown and the shape of the equity curve.

The three big distortions

Overfitting: optimising parameters until history looks perfect models noise instead of signal – the strategy collapses live. Antidotes: few parameters, in-sample/out-of-sample separation, walk-forward analysis. Look-ahead bias: the simulation uses information not yet available at decision time (e.g. the closing price for an entry within the same candle). Survivorship bias: datasets without failed, delisted assets systematically overstate historical returns.

Putting validity into perspective

A backtest is an estimate under the assumption that market behaviour repeats structurally – regime changes (rate turns, volatility shocks) regularly break this assumption. Therefore: the sample must be large enough (rule of thumb: several hundred trades across different market phases), costs must be set conservatively, and between backtest and live use belongs a forward test on unseen data or in a simulator.

Why it matters for traders

Backtesting replaces opinion with measurement: it shows whether a setup had any historical edge at all and delivers the metrics for position-sizing and drawdown planning. In the GlanWick simulator strategies can be played through on historical data by way of example and the metrics evaluated automatically – GlanWick is a training and simulation tool and not a prop firm itself.

Execution Example

A trader backtests a moving-average crossover over 5 years of price data on a simulated $100,000 account: 240 trades, 42% win rate, average winner +$1,900, average loser −$900, fees and slippage set at $30 per trade.

Expectancy before costs: 0.42 × 1,900 − 0.58 × 900 = 798 − 522 = +$276 per trade.
Costs: −$30 per trade → net expectancy +$246; over 240 trades ≈$59,000 simulated profit.
Profit factor: (101 winners × 1,900) ÷ (139 losers × 900 + 7,200 costs) = 191,900 ÷ 132,300 ≈ 1.45.
Robustness check: compare in-sample (years 1–3) vs. out-of-sample (years 4–5) – if expectancy collapses to near zero out-of-sample, the parameter choice was over-optimised and the test is repeated with simpler rules.

Execution Risk & Errors

Optimising parameters until history looks perfect (overfitting)

Omitting fees, spread and slippage and flattering the expectancy

Building in look-ahead bias, e.g. entries based on candles not yet closed

Drawing statistical conclusions from 20–30 trades

Treating backtest results as a forecast instead of a historical estimate

Frequently Asked

How many trades does a meaningful backtest need?

As a rule of thumb several hundred trades across different market phases. Small samples make luck look like edge – the dispersion of results is then larger than the measured advantage.

What is the difference between backtest and forward test?

The backtest simulates on known past data, the forward test runs on new, unseen data – e.g. in a simulator or live at minimum size. Only both together give a reliable picture.

What does out-of-sample testing mean?

The data is split: the strategy is developed on one part (in-sample) and validated on the held-back part (out-of-sample). If performance collapses there, the strategy is over-optimised.

Why do backtested strategies fail live?

Most common reasons: overfitting, underestimated costs, market regime changes and execution differences between simulation and reality – such as slippage in fast phases.