Thread by @ReformedTrader

1/ Adaptive Asset Allocation: A Primer (Butler, Philbrick, Gordillo, Varadi)

"Estimates of parameters for portfolio optimization based on long-term observed average values are inferior to estimates based on observations over much shorter time frames."

papers.ssrn.com/sol3/papers.cf…

2/ This very readable article walks through the reasons why the inputs usually used for portfolio optimization are not reliable enough, especially for the 3-5 year emotional horizons most investors have.

3/ Returns, volatilities, and correlations can deviate from their historical averages by large amounts.

More recent observations tend to be better predictors; this is why momentum and volatility targeting have historically worked. The choice of lookback window matters.

4/ The authors start with ten major global asset classes; real ETF and mutual fund returns were used when possible.

Moving from equal weighting to inverse volatility weighting (60-day measurement window) improves volatility-adjusted returns, drawdowns, and return consistency.

5/ Choosing the top 5 (of 10) asset classes by 6-month momentum results in further improvements (with inverse volatility weighting > equal weighting).

Using recent measurements for returns (momentum) and volatility seems to work, but we haven't looked at correlations yet.

6/ A minimum variance portfolio formed using the same 5 asset classes takes correlations into account, resulting in even better metrics.

(Note to self: Minimizing variance could work for value as well, but maximizing Sharpe could be a problem because value stocks mean revert.)

7/ The previous risk-adjusted returns assumed a zero return on cash and no transaction costs.

After a 1% management fee, trade frictions that scale down linearly from 2.5% in 1995 to 0.5% in 2013, and the 10-year Treasury yield, the strategy isn't magical but still looks good.