Lib Freqtrade

freqtrade

A live-first retail crypto bot that ships the best bias-detection tooling in the entire domain — and an optimistic backtester whose assumptions it is honest enough to publish in full.

GPL-3.0 is the strongest copyleft in this catalog. A strategy file freqtrade loads at runtime is one thing; shipping a product built on freqtrade is another — distribution obliges you to release source under the same terms. Internal and personal use is unaffected. If the deliverable is proprietary, use jesse (MIT) or build on ccxt directly.

The trap that costs you money

Zero slippage. Always. Every order fills at the requested price provided that price sits inside the candle's high/low. Real fees are charged (pulled from ccxt); market impact is exactly zero. On thin alts this is the single largest source of phantom P&L.

Its twin: stoploss exits fill exactly at the stoploss price, even when the candle's low was lower. A gap straight through your stop costs you nothing in backtest and a great deal live.

The rest of the documented assumptions

freqtrade publishes its own backtest assumptions, which is why this list exists at all:

• Within a candle, the high is assumed to occur before the low — which is what makes trailing stops look good. You cannot know intrabar ordering from OHLC; every bar engine guesses, and freqtrade guesses in the direction that flatters trailing stops.
• Exit signals take priority over stoploss, because exits are assumed to trigger at the next candle's open. A candle that would have stopped you out mid-bar instead exits on your signal.
• Survivorship via live pair lists. VolumePairList and friends build the universe from today's listings, so backtesting 2021 on today's pairs misses every token that went to zero. Use a StaticPairList rebuilt from archived fetch_markets() snapshots, or call the result an upper bound.
• The hyperopt maximum is not a result. A 2,000-evaluation hyperopt is 2,000 trials for Deflated-Sharpe purposes.
• stoploss_on_exchange is off by default. On, the stop lives at the exchange and survives your process dying, your VPS rebooting or your network dropping. It is the most important single line in retail crypto risk management, and you must opt in.

Other documented behaviour: entries fill at the open unless custom pricing applies; ROI exits compare against the candle high but never price below the low; stoploss is evaluated before ROI; order is exit signal → stoploss → ROI → trailing stoploss. The project's own closing statement is that backtesting will never replace dry-run mode, because intrabar ordering is unknowable.

Every bar engine makes assumptions of this shape. freqtrade is simply the only one that lists them. Treat that list as the template for interrogating any other engine.

The two bias detectors — worth running from another framework

freqtrade lookahead-analysis runs a baseline backtest, then re-runs it over progressively sliced windows and compares dataframe columns. Divergence means the indicator saw the future. It catches shift(-N), raw .iloc[] indexing, unrolled full-series aggregations (.mean(), .max()) and badly controlled loops. Stated limits: it only checks triggered signals (untriggered leaks escape as false negatives), false-positives on limit orders with custom pricing callbacks, may falsely flag FreqAI target indicators, and is useless for rarely-signalling strategies.

freqtrade recursive-analysis varies startup_candle_count and reports each indicator's last-row variance against the base calculation. - = converged, nan% = insufficient data, a large percentage = raise your startup candles. It is the only off-the-shelf tool that measures the recursive warm-up problem: EMA/RSI/ADX converge to different values depending on how much history preceded them, and a backtest sees 5,000 candles where live sees the ~1,000 the exchange returns in one call. Its stated mechanism — backtesting loads the whole dataframe into memory while live processes candles sequentially — describes vectorbt, PyBroker and backtesting.py equally well.

Minimal correct call

{
  "dry_run": true,                 // default in the shipped example config — keep it until deployed
  "dry_run_wallet": 1000,
  "stake_currency": "USDT",
  "trading_mode": "spot",
  "order_types": {
    "stoploss_on_exchange": true,  // OFF by default; turn it on
    "stoploss_on_exchange_interval": 60
  },
  "pairlists": [                   // NOT VolumePairList — that is survivorship bias
    { "method": "StaticPairList" }
  ],
  "exchange": { "ccxt_config": { "enableRateLimit": true } }
}

freqtrade backtesting        --strategy MyStrat --timerange 20210101-20240101
freqtrade lookahead-analysis --strategy MyStrat   # run BEFORE believing any result
freqtrade recursive-analysis --strategy MyStrat

Live keys must carry trade scope only, never withdraw; dry_run: true is not proof you are safe.

See also

• ../../../fin-crypto/skills/crypto-data-and-execution/SKILL.md — survivorship, funding, 365 days
• ../../../fin-crypto/skills/crypto-data-and-execution/references/freqtrade.md — reference card
• ../../../fin-crypto/skills/crypto-data-and-execution/references/ccxt.md — the exchange layer
• ../../../fin-core/skills/backtesting-engines/references/execution-realism.md — fill realism

Where this sits

This file is the deep dive on one library and assumes the choice is already made. For which library to pick, how it compares with the alternatives, and the traps that span several of them, the entry point is the domain skill backtesting-engines (../../../fin-core/skills/backtesting-engines/SKILL.md).