Why backtest
Before investing real money, find out how your strategy would have worked on historical data. If it was unprofitable in the past, it will likely be unprofitable in the future.
Source — Binance Historical Data
Binance provides free OHLCV (Open/High/Low/Close/Volume) data for the entire time period. You can access it via REST or bulk download.
👤 Prompt
Create a backtester in your grid-bot:
1. data_loader.py: download OHLCV ETHUSDT 1-minute data for the last 60 days via the Binance API. Save it in parquet format.
2. simulator.py:
- Load the data.
- Initialize the grid with the same parameters as your live config.
- Go through each minute:
- If the price touches the buy level → simulate a fill, add a sell order at the next level.
- If the sell level is reached → fill, buy at the next level.
- Track: filled orders, position size, USD value.
3. report.py:
- Total profit/loss over 60 days.
- Daily breakdown.
- Maximum drawdown.
- Win rate (filled grid trades).
- If there was a breakout — what happened.
Run the backtest. Show the report + equity curve graph.
What to look for in the results
| Metric | Good indicator |
|---|
| Total return over 60 days | +5-15% |
| Max drawdown | <-10% |
| Daily P&L distribution | Most days are profitable |
| Breakout events | <3 in 60 days |
| Sharpe ratio | >1.5 (excellent >2.5) |
Walk-forward optimization
If you want to optimize parameters (step, range) — don't do it on the entire dataset. Otherwise, you'll overfit.
👤 Prompt
Add walk-forward.py to your backtester:
1. Split historical data into 70/30 (train/test).
2. On the train data — try different step values (0.5, 1, 1.5, 2, 3%) and range values (±3%, ±5%, ±8%, ±10%).
3. Choose the best parameters based on the Sharpe ratio.
4. On the test data — check how they work (out-of-sample).
5. If the test results are <30% worse than the train results — it means overfitting, don't use them.
Show the top 5 combinations + their out-of-sample performance.
Reality check
Backtesting shows an idealized picture. A real bot can be 20-40% worse:
- Slippage — your orders are executed at a non-precise price (especially market orders).
- Fill delays — in backtesting, fills are instantaneous, in reality, they take seconds or minutes.
- Network outages — the bot may not respond to rapid movements.
- Funding rates (on futures) — they take away part of the profit.
So if backtesting showed +12% — expect 6-8% in reality. If backtesting showed -5% — RUN, reality will be worse.
💡 Pre-flight checklist
Before mainnet:
✅ Backtest for 60+ days — net positive
✅ Walk-forward — out-of-sample not deg more than 30%
✅ Drawdown <10%
✅ Sharpe >1.5
If all 4 — launch. If any of them are not met — the parameters are not suitable.
🎯 Key takeaway
Backtest for 60 days. Walk-forward to avoid overfitting. Realistic expectation — 70% of backtest results. In the next lesson — risk management and breakout protection.