Choosing Historical Data Windows for Portfolio Optimization
Summary
Portfolio optimization uses historical returns to estimate a variance-covariance matrix, which supplies estimates of asset volatilities and covariances. The amount of history needed depends partly on return frequency: intraday and daily data may also require adjustment for market noise, while monthly observations produce fewer data points over the same span. The response gives a rough example that a year may suffice for daily returns but may be inadequate for monthly returns.
More observations can make estimates more robust, but older data may describe relationships that have since changed. Selecting a window therefore means making a judgment about which past conditions are relevant to the future, including whether to emphasize recent conditions or periods of sharp moves. The discussion also mentions overriding historical estimates with forecasts and considering Black-Litterman as one approach to incorporating views. It offers no universal window length or empirical comparison; the appropriate choice depends on sampling frequency, changing correlations, and the market regime relevant to the portfolio.
Key ideas
- Portfolio optimization commonly estimates a variance-covariance matrix from historical returns.
- The observation frequency affects how much calendar history is needed to obtain a useful sample.
- Longer histories can stabilize estimates, but may include relationships that no longer hold.
- Window selection is a judgment about which historical market conditions may be relevant going forward.
- Forecast views can supplement historical estimates, including through approaches such as Black-Litterman.
Tags
Full text
# How many years of historical data is require for Portfolio Optimization? # How many years of historical data is require for Portfolio Optimization? I would like to know about below questions. ``` How many years of historical data is require for Portfolio Optimization in R programming. ``` Thanks Atul ## Answer by dnl (score 2, accepted) https://quant.stackexchange.com/a/26057 Speaking about portfolio optimization in general, historical data is required to estimate the variance-covariance matrix that gives you the volatilities of the stock returns and the covariances thereof. The estimation depends on the periodicity of your returns, i.e. whether you are using intraday, daily, weekly, or monthly returns. (For intraday/daily returns you would also need a noise adjustment.) Theoretically, you can use any length of time series to estimate the matrix. The more data you use to estimate the matrix the more robust it will be. But the correlation relationship between the returns may change over time so it can be feasible to not use all available data. The required time series length also depends on the mentioned periodicity of your returns. For daily returns one year (250 trading days) may be enough, for monthly returns one year may be too few. ## Answer by horseless (score 1) https://quant.stackexchange.com/a/26081 I agree with everything dnl said. This time period choice can get very involved. You are explicitly choosing an historical period that you think will be repeated in the future. Unfortunately, history does not repeat exactly. There is an old saying, all models are wrong, but some are useful. There is no right answer for the period. Should you include just a recent past or include periods of sharp market moves? Additionally, there are those that just over-ride the correlation matrix and returns numbers because they explicitly believe they can forecast better than using a previous period. You might want to explore the Black-Litterman model to see how some try to improve on the efficient frontier model. If I were doing this, I might look at the correlations estimated only during a rising interest rate environment.
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.