Скачать или смотреть The Biggest Algorithmic/High Frequency Trading Backtesting Mistakes You Can Make in Python, C+, Ect

Backtesting is testing trading strategies against historical data.

It serves as a research tool for traders and analysts to stress test their current strategies, find new strategies, or find which factor contributes the most to a strategy’s success.

Most backtesting is done with code, using libraries from languages like Python and R and hence, you can pretty much backtest anything if you have the coding skills to create it.

There are also some no-code or low-code backtesting solutions like TradeStation or MultiCharts, but they tend not to be flexible enough for professionals.

An example of backtesting would be to download a decade’s worth of market data, let’s say 2007 to 2017, and test a trading strategy you’re curious about. You need to be able to put the strategy into a strict quantitative format, like “buy when the 50-day moving average crosses above the 200-day moving average,” instead of “buy a bull flag.”

After the computer is done running the test, you get results like an equity curve, the list of trades, and some metrics like Sharpe ratio and max drawdown.

Here’s an example of a “tearsheet” of analytics generated when you run a backtest on the Zipline Trader library, an updated version of now-defunct Quantopian’s Zipline library:


Backtesting can feel like magic.

Download a bunch of market data, run some parameters against it, tweak them a bit until the test shows a smooth equity curve and presto, you have a profitable trading system that will print money. It can seem like the only barriers are learning to code and understanding which indicators and parameters will yield the best results.

But doing that is simply data mining and using this method of backtesting has little utility because your tests will have no predictive power. It represents a mistake even the most sophisticated researchers repeatedly make: mistaking correlation for causation. I’ll demonstrate this with an example:

You’ve run thousands of tests using machine learning to find the most profitable trading system.

The best test shows that the optimal strategy is to buy XYZ stock at 10:53 AM on a Tuesday when on the previous day, the price advanced at least 1.04% and the RSI is at 38 or more. Obviously, this means nothing. If I told you this was my trading system, you’d laugh at me for having such a pointless system that doesn’t exploit any real imbalances in the market.

All this test did is find the perfect parameters to fit to that exact historical data.

While data mining when backtesting is rarely this blatantly pointless, you have to constantly check every bias you introduce into a test.

Here’s a few of the most common mistakes that are crucial to avoid if you don’t want to incinerate money when trading with backtested strategies.

Look-Ahead Bias
It’s 2021 and the market has more or less gone straight up for the last decade, save for a huge crash in 2020.

Knowing this, you can go create a backtest against the historical data to reflect this. For example, you might say, buy the S&P 500 on 5x margin, but sell when the price drops 5% or more in one day. You know that this system will kill it, because it avoids most of the 2020 market crash, while reaping huge gains from the rest of the bull market.

A test like this will look really nice and can make you wish you only had the system in 2010, because you’d be rich right now.

But you’re basically cheating.

It’s like watching a replay of a football game, and then concluding that the best way to win that game was to run the ball on 4th & goal, because passing the ball failed.

You’re just looking at what actually happened in the past, and pretending like you knew the right answer all along. It might be a fun exercise, but it won’t have any future predictive power.

Of course, look-ahead bias creeps into our backtests in more subtle ways. Using the same bull market example, momentum strategies on growth stocks have worked excellently, and traders who used those strategies over the last decade won big.

This can easily lead you to believe that momentum strategies are inherently superior and tailor your entire style to what is working in recent backtests.

Over Fitting or Over Optimization
This mistake goes hand-in-hand with look-ahead bias. It’s the act of continually iterating on a backtest and using previous backtest results to inform the new ones. While this can be a valid practice used by many professionals, it’s very easy to get wrong.

Let’s say you’re using a simple RSI mean reversion strategy.

You buy when RSI is below 20 and close the position when RSI is above 60. The backtest results look pretty good, but when you analyze a few individual trades you find that the biggest losers result in you buying too early; in other words, you think a lower RSI value might result in a better test. So you try testing buying below 15 and selling above 60.