It doesn’t matter how well your model fits the data. What matters is how well it performs out-of-sample (OOS). I’ve just uploaded a research note documenting the OOS performance of my retail and wholesale strategies for trading the term structure of expected volatility. If you want the details, go there. I thought I’d upload the pictures here for the tl;dr types.
Basically, I have managed to isolate a very strong signal of intermediary risk appetite at arbitrary frequencies. Since fluctuations in dealer balance sheet capacity drive fluctuations in asset prices, this is a highly lucrative feature. In a sense, dealers live on the short-end of the volatility curve. They aggressively respond to innovations to systematic volatility by rescaling their balance sheets. The price of systematic volatility is the dual of dealer risk appetite. When risk appetite is high and balance sheet growth is strong, the price of vol collapses; when risk appetite is low and balance sheet growth is weak or negative, the price of vol rises. What I have managed to do is read off fluctuations in risk appetite from VIX futures. This allows me to predict a risk-off tomorrow after observing the signal today.
In order to pin down my predictive model, I need to tune a hyperparameter. We use 5-fold OOS cross-validation (CV) with Prado’s log loss as our loss function since our regime switching strategy is especially exposed to bad predictions with high confidence.
The feature is uncannily good at predicting risk-offs on the next day. The major spikes correspond to well-known events. The first big spike corresponds to the “China panic” on August 24, 2015. The second one, corresponds to the dramatic return of volatility on February 5, 2018. Our predictive model called both of them. It also called the doldrums towards the end of 2018, and, of course, the dramatic revival of systematic volatility associated with the Coronavirus pandemic that got going on February 24, 2020 and is still underway. My model told me last night that there was a 59.4 percent chance that there would be risk-off today.
On all of these and a few others, the predicted probability of risk-off exceeded 50 percent. But what is the appropriate threshold to call a risk-off? When should we move from being short the term spread, ie selling vol while hedging with medium-term vol, to being long? The probability cutoff is location hyperparameter than we tune, again with 5-fold CV, to achieve the highest Sharpe ratio OOS.
Now that we have pinned down our hyperparameters, we are ready to evaluate our trading strategies against the benchmarks. The next few figures display the in-sample and OOS Sharpe ratios, max drawdowns, return on max drawdowns, and cumulative returns since Jan 1, 2014.
Finally, Table 1 displays the returns statistics. The numbers speak for themselves so I won’t rub it in. Let me know if you want access to the secret sauce.