How to backtest a trading bot, then prove it with paper and forward tests

BotIndex editorial team · Updated 11 Nov 2026 · Editorial policy

How to backtest a trading bot: the four stages

Knowing how to backtest a trading bot is half the job. The other half is knowing what a backtest cannot tell you. Treat testing as four stages, each cheaper to fail than the next:

People use “paper trading” and “forward testing” loosely. In this guide, paper trading is the mechanics check and forward testing is the honest scorecard: you freeze the settings and let new data judge them. Freqtrade’s backtesting docs put it bluntly: backtesting “will never replace” its dry-run mode, and “only forward-testing (dry-mode) can really confirm a strategy”.

If you have not built the bot yet, start with how to build a trading bot. This page goes deeper on the testing step.

What a trading bot backtest must include

A backtest is only as honest as its cost model. Before you look at the profit line, check that it includes:

• Trading fees on both sides. Every round trip pays an entry fee and an exit fee. Freqtrade applies its fee “twice (once for trade entry, and once for trade exit)” and by default uses the exchange’s listed fee. Use the fee you will actually pay, including whether your orders are maker or taker.
• Slippage. Many backtesters assume perfect fills. Freqtrade’s docs state that orders fill “at the requested price (no slippage)” as long as the price is inside the candle, and that stop-losses exit exactly at the stop price “even if low was lower”. Real markets are not that kind, so add a slippage allowance yourself.
• Funding on futures. Perpetual futures charge or pay funding at set intervals. A strategy that holds positions for days can lose its edge to funding alone, and our HTX review notes that its futures grid pays normal trading fees plus funding.
• Realistic fills and sizes. Exchange minimum order sizes, partial fills on thin pairs, and the fact that a candle’s high and low happened in an order you cannot see.
• Subscription costs. If the bot runs on a paid platform, the monthly fee is part of the cost. Our bot fee calculator adds subscription and trading fees at your volume.

Worked example: how fees change a backtest result

Example with made-up round numbers, not a real strategy. A spot bot trades a fixed 1,000 USDT per trade. Over a year of history it makes 150 round trips, and its average gross gain is 0.25% per trade. Profits are not compounded.

• Gross result, no costs: 150 × 0.25% × 1,000 = +375 USDT. This is what an untuned backtest might show.
• With 0.10% taker fees on each side (the spot rate in our Bybit review): 0.20% per round trip, or 2 USDT × 150 = 300 USDT in fees. Net +75 USDT.
• Add 0.05% slippage on each side: another 0.10% per round trip, or 150 USDT. Net −75 USDT. The strategy now loses money.
• Same trades at 0.05% fees on each side (the spot rate in our Pionex review), still with slippage: fees 150 USDT, slippage 150 USDT. Net +75 USDT.

The signal never changed. Only the cost model did, and the verdict flipped twice. High-frequency strategies with small average gains, such as tight grids, are the most exposed. Our grid profit calculator shows the same effect for a single grid cycle.

Common backtesting traps

Overfitting

Overfitting means tuning settings until they fit past noise. Try enough combinations of periods, thresholds and stops, and one will look brilliant by chance. Bailey, Borwein, López de Prado and Zhu’s paper on the probability of backtest overfitting argues that standard safeguards such as a hold-out sample “tend to be unreliable” for investment backtests, and proposes a way to estimate how likely it is that a chosen strategy is overfit. Freqtrade’s docs make the same point in plain terms: when many strategies are tested, “some will score a low p-value by chance alone”.

Practical defences: keep the number of parameters small, record how many variations you tried, test on a period you never tuned on, and distrust a result that collapses when you nudge a setting slightly.

Look-ahead bias

Look-ahead bias is using information the bot could not have had at the time. Freqtrade’s lookahead analysis page says that if your indicators or signals “look into future candles, this will falsify your backtest”, and it ships a command that re-runs backtests to detect it. Typical causes: using the current candle’s close to enter at that same candle’s open, or pulling higher-timeframe data before that candle has closed.

On TradingView, the related problem is repainting. Its repainting docs explain that historical bars are always confirmed while a live bar keeps changing, so a signal can appear in the backtest that would have vanished live. Our TradingView trading bot guide covers how to set alerts to fire on bar close.

Survivorship bias

If you backtest only on coins that are popular today, you skip the ones that were delisted or collapsed. Your test universe has been filtered by the outcome, which makes any strategy that buys dips look safer than it was. Use the pairs that were actually tradable at each point in the test, or at least include coins that later failed.

Choosing a friendly period

A grid or DCA bot tested only on a sideways or rising market says little about a crash. Include at least one sharp fall, one strong trend and one quiet range. If you use leverage, check the worst drawdown against your liquidation price, not just the final profit.

Which platforms offer backtesting or paper trading

From our review data. “Paper” here means simulated or demo trading.

Outside our reviews, OKX’s signal bot FAQ describes a demo trading environment with its own webhook URL, and Freqtrade is free open-source software with backtesting, lookahead analysis and dry-run built in.

Be careful with backtests a provider shows you. The SEC’s Investor.gov bulletin on performance claims says “back-tested performance is hypothetical and does not reflect actual performance”. A marketplace bot’s backtest is a sales claim until you reproduce it, which is one reason we cover whether crypto trading bots are profitable separately.

Trading bot backtesting checklist

1. Write the rules down before you open any chart: entry, exit, size, stop.
2. Split your data: tune on one period, test once on a later period you never touched.
3. Model costs: both-side fees at your real tier, a slippage allowance, funding for futures, and the subscription.
4. Check for look-ahead: signals use only closed candles; run a lookahead check if your tool has one.
5. Use a fair universe: include pairs that were delisted or crashed, not only today’s winners.
6. Cover bad markets: at least one crash, one trend and one range.
7. Read the drawdown, not just the profit. Size it with our position size calculator and the rules in trading risk management.
8. Count your attempts. If you tried dozens of variations, assume the best one is partly luck.
9. Paper trade, then forward test with frozen settings, and compare trade by trade with the backtest.
10. Go live small, and stop if live results drift well below the forward test.

FAQ

Sources

• Bailey, Borwein, López de Prado, Zhu — The Probability of Backtest Overfitting (WMU ScholarWorks) ↗checked 11 Nov 2026
• Freqtrade — Backtesting ↗checked 11 Nov 2026
• Freqtrade — Lookahead analysis ↗checked 11 Nov 2026
• TradingView — Pine Script docs: Repainting ↗checked 11 Nov 2026
• Investor.gov (SEC) — Investor Bulletin: Performance Claims ↗checked 11 Nov 2026
• OKX — Trading Signal Bot FAQs (demo trading) ↗checked 11 Nov 2026

Not financial advice. Trading bots do not remove market risk.