Backtrader Python Backtesting

What it is

Backtrader is an open source Python library for backtesting trading ideas and, through broker integrations, running them live. It is a library, not an application: you write Python code that uses it, so the quality of the result depends on your code and your data as much as on the library. Being open source, it is free to read and change, and its maintenance status and the state of its integrations are things to check on the project's own page, because open source projects can slow down or stop.

The vocabulary is worth knowing. A central engine, usually called Cerebro, holds data feeds, strategies, a broker simulation and analysers. A strategy is a class with a method called once per bar. Indicators are objects the strategy creates once and reads each bar. This page explains how it works in general terms. We have not tested it for this page. Check the vendor's current terms.

How a bot connects here

The connection style is your own Python code using a library, with optional broker integrations for live use.

• Data feeds. You load data from a file, a data frame or a vendor. The library does not check that it is correct.
• The strategy method. Called once per bar. Your logic decides to buy, sell or wait.
• The simulated broker. It applies a cash balance, commission and, if you set it, slippage. By default the settings are plain, so set them to match reality.
• Order timing. In libraries of this kind a market order sent on one bar is usually filled on the next bar, at its open, by default. Read the documentation for the current rule and test it.
• Live use. Through integrations that you or others maintain. Check they are current before relying on them.

Worked example: warm up eats data

Every indicator needs a number of bars before it can produce a value. A strategy built on a 200 bar average cannot trade until bar 200. If the test data is short, much of it is spent warming up, and the remaining trades may be too few to mean anything. Numbers are made up and checked in node.

With 1,000 bars and a 200 bar warm up, 800 bars (80 percent) remain. With a year of daily data, about 252 bars, a 200 bar warm up leaves only 52 bars, 21 percent, and a test over 52 bars cannot tell skill from luck. Cutting the longest indicator to 50 bars leaves 202. The lesson is to match the length of the rule to the length of the data, and to check the number of trades produced, not only the profit line. See how to backtest and the Python trading bot page.

What to verify before trusting a backtest here

These checks apply to any platform of this kind. The answers come from the vendor's current documentation and from your own tests, not from this page.

How bots break on this kind of platform

• Stale integration. A live link that no longer matches the broker's current API.
• Optimistic fills. Stops and limits filled at the best price in a bar.
• Logic written for the test. Code that works on a complete data frame but not on a stream.
• No state on restart. A live bot that forgets its positions.
• Unclear timing. Look-ahead by reading data you should not have. See look-ahead bias.

A safe order of work

Backtest with honest costs, forward test on new data, run on a simulated account for weeks, then trade a very small size and compare each step with the one before. See forward testing and paper trading.

What this page does not tell you

It does not tell you whether the product is suitable, what it costs, or how it behaves today. Those change and belong to the vendor. It does not give a strategy to trade. This page explains how it works in general terms. We have not tested it for this page. Check the vendor's current terms.