Backtest before you bet: pull years of free candles from any exchange

Every strategy idea feels brilliant at 2 a.m. The cheapest way to find out whether it actually is: run it against history before you run it against your savings. Exchanges give away that history for free — candles (OHLCV: open, high, low, close, volume) over public endpoints, no API key required — and CCXT exposes all of them through one call: fetchOHLCV.

The catch, as covered in mistake #5: one call returns one page, capped per exchange (commonly 500–1000 candles). Downloading years of data means looping. Here's the loop done right.

Downloading full history

Request from a start time, advance since past the last candle received, stop when a page comes back short:

import ccxt from 'ccxt';

const exchange = new ccxt.binance();
const symbol = 'BTC/USDT';
const timeframe = '1d';
let since = exchange.parse8601('2020-01-01T00:00:00Z');
const all = [];

while (true) {
    const candles = await exchange.fetchOHLCV(symbol, timeframe, since, 1000);
    if (candles.length === 0) {
        break;
    }
    all.push(...candles);
    since = candles[candles.length - 1][0] + 1;   // 1 ms past the last candle
    if (candles.length < 1000) {
        break;                                     // short page = caught up
    }
}
console.log(`fetched ${all.length} candles`);
await exchange.close();

The built-in rate limiter paces the loop for you. Daily candles since 2020 arrive in a few pages; minute candles for the same period are ~3 million rows — still perfectly doable, just let it run and write to disk as you go instead of holding everything in memory. Save to CSV/Parquet once and iterate on your strategy offline; re-downloading history on every run is how people get rate-limit bans while "just backtesting."

The four traps that make bad strategies look good

Downloading candles is the easy half. Most homegrown backtests are broken in one of four ways — always in the flattering direction:

1. Lookahead bias. Your signal uses information that wasn't available at trade time — the classic version is acting on a candle's close during that candle, or computing an indicator over a window that includes the current bar. Rule: signals computed on candle N execute at the open of candle N+1, at best.

2. No fees, no slippage. Add taker fees (market['taker'], commonly ~0.1%) on every fill, and a slippage haircut on top. High-frequency strategies that look great at zero cost usually die at 2×fees. If your edge per trade is smaller than round-trip costs, you don't have an edge — you have a fee-generation machine.

3. Unrealistic fills. A backtest that fills limit orders whenever price touches your level is lying to you — in reality you're at the back of the queue, and the fills you do get skew toward the times the market blows through your price (adverse selection, the same enemy from the market-making post).

4. Survivorship and overfitting. Testing only on BTC/ETH — coins you already know survived — inflates results; so does tuning parameters until history looks perfect. Hold out a time period you never touch until the end, and treat a strategy that only works with rsi_period = 13 (but not 12 or 14) as noise, not signal.

From candles to a verdict

Keep the pipeline honest and boring: download once with the loop above → store to disk → compute signals bar-by-bar as if live → next-bar execution with fees and slippage → compare against buy-and-hold, because a bot that underperforms doing nothing is expensive entertainment. If it survives all that and an out-of-sample period, promote it to the testnet with setSandboxMode(true) — paper trading against live prices is the final exam that history can't grade.

From there, the first-bot tutorial covers the execution side, and WebSockets the live data feed.