How Duplicate Instruments Can Bias Backtests

Summary

The document considers how duplicated or apparently similar securities can affect strategy research. Its responses distinguish true duplicates from instruments that merely share an issuer: Alphabet’s two share classes, for example, have different voting rights and slightly different price histories. Testing both series might produce similar, but not identical, results for a system applied to each.

The broader answer is that the effect depends on the task. In asset allocation, repeated observations can alter portfolio weights; in classifier or regression work, they can change the regressors and distort the fitted model. Such distortions belong to the wider family of sampling bias in backtesting. The discussion gives no quantitative experiment, cleaning procedure, or measure of the resulting bias, and it does not establish that all datasets contain duplicates. Researchers should identify whether records represent duplicate observations or genuinely distinct securities before deciding how to handle them.

Key ideas

• Repeated data can alter asset allocation weights or the inputs to statistical models.
• The impact of duplicates depends on the research task and how observations are used.
• Securities from the same issuer may represent distinct share classes with different price histories.
• Duplicate observations are a form of sampling bias that can distort backtest conclusions.

Tags

# How does duplicate data affect backtesting?


# How does duplicate data affect backtesting?












Before developing a trading strategy, one should clean and preprocess his data set. It is very common that a data set contains huge number of duplicates.

> Question: How does duplicate data affect backtesting?

## Answer by Garima Gulati (score 4)

https://quant.stackexchange.com/a/47188

As clarified by you in the comments, I haven't come across duplicate tickers at least not in Bloomberg or the Indian exchange websites. Example of GOOG and GOOGL mentioned in the comment, represent Alphabet Inc Class C and Alphabet Inc Class A respectively. Class A shareholders enjoy voting rights whereas Class C shareholders don't (see here).

Hence there is a slight difference in their price series.

Ideally, if you are testing any system on both the price series, you should get similar results. The numbers won't be exactly the same but they will be similar.

## Answer by Vitomir (score 3)

https://quant.stackexchange.com/a/47185

It depends what you are doing: asset allocation, classifier/regression modelling, etc. In any case can shift the weights of the allocation, change the regressors, etc. Nevertheless, it belongs to the family of backtesting sampling bias.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.