Model ETFs Individually and as a Group to Compare Returns
Summary
For a relatively homogeneous set of US large-cap ETFs, the answer recommends studying both individual return models and a combined group model. A group series can act as an index for comparing each ETF’s performance and estimating individual or group relationships to the broader market, such as subgroup beta.
The choice also depends on how predictions will be traded: a combined forecast does not make the instruments interchangeable, since trades are placed in specific securities unless the group itself is represented by a fund. The response considers a dataset of roughly 15 years per ETF and argues that this scale is manageable with common data tools. It does not compare model accuracy or prescribe a particular modeling technique. A key practical caveat is to adjust historical prices for corporate actions, including splits and dividends, before estimating returns.
Key ideas
- Modeling ETFs individually and as a group can provide complementary comparisons.
- A group series can serve as a benchmark for individual performance and beta analysis.
- A forecast across a group still has to be translated into trades in specific instruments.
- The example dataset is described as manageable with standard data tools.
- Historical prices need appropriate split and dividend adjustments.
Tags
Full text
# Modeling EOD ETFs price returns together or individually? # Modeling EOD ETFs price returns together or individually? Let's say you want to model the next day price returns for a set of US equities large cap ETFs (a relatively homogenous group). Would you model all the ETFs as a single, 15 years data set, or each ETF individually? Each ETF would have around 15*252 = 3780 unique data points. I know, it's a general question that is relative to the predictors used, the strategy goal and other things. But what are the plus and minuses of each approach? ## Answer by Brad S (score 1) https://quant.stackexchange.com/a/20849 Model them individualy and as a group. When you model them as a group you are essentially building a stock index that you can compare the performance of individual stocks to and can then calculate a subgroup beta for each stock. You can also calculate a beta coefficient for the group as a whole to the wider market. Since I assume that you are modeling them with the intention of trading on the prediction output, you need to have a sense of the variance of each stock due to the discrete nature of placing trades. You cannot enter a single trade that buys the whole group, unless the group is already a fund of funds ETF. To build an index you neeed to include all of the data points anyway. 3780 unique points is simple in post 2007 Excel 64 bit model. A solution with a SQL database in any popular programing language will also make 3780 points seem trivial. The challenge will be to ensure you normalize price movements for stock splits, dividends, etc over such a long time period.
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.