BOTW 8: Rolling with AAII

Rolling regression on AAII signals improves strategy results meaningfully

We revisit our analysis of the American Association of Individual Investors (AAII) Survey for our backtest of the week. Recall in our last post, we discussed how using the AAII survey as a signal showed some promise when coupled with a little bit of market knowledge. But it wasn't that great either. We conjectured that using it as a conditioning, rather than directive signal, might prove useful. But we were encouraged by the results simply because we found a model that seemed pretty stable and with reasonable logic. That said, we decided we weren't done seeing if we could wring a bit more juice from this lemon. And indeed it might be sumo orange instead!

We've used rolling regressions for other backtests. Frequently, they've improved results nicely. The main intuition here is that the static model built from a train set common in machine learning just doesn't apply to time series data. The data, outlook, even the regime are changing often daily. So, like Keynes, we should ask our model what it does when the facts change. Most ML models don't do anything. But rolling regression and walk-forward models do. So let's use them.

Recall in our last backtest we used a train/valid/test split. We trained the model using the bullish, bearish, bull-bear spread, and initial reaction features to predict the next week's return prior to the next survey release. For this backtest, we'll use the same split, but the features will be the bearish, 8-week moving average bullish, and 8-week moving average bull-bear spread sentiment indicators as well as the initial reaction. We'll then create multiple rolling regressions models by iterating through four different lookback windows—13, 26, 52, and 104 weeks—and three different thresholds—negative one, zero, and positive one. You can see the results below. We group the outperformance of the strategy relative to the benchmark by window and threshold. It appears the shorter window lengths perform better. Due to quicker updating or fast markets? Something to examine in another post.