Regime-Based Investing, ep 2

Man AHL dropped a new paper on regime-based investing (link). Think of it as the overachiever’s version of lazy portfolios. Instead of building a portfolio that’s resilient across economic regimes—like the Permanent Portfolio—this approach tries to front-run reality. If we’re in stagflation, why hold anything but cash? Or managed futures, hehehe. If growth is up and inflation is asleep, just ride stocks.

The pitch? Diversification, but make it dynamic. Instead of spreading bets across assets, you rotate into whatever should be working right now. Diversification without the apology tour.

Sounds amazing. Why doesn’t everyone do it?

Because, like most things that sound too good to be true, it usually is. Execution is the graveyard. Turnover eats returns. The “right” asset might still go down. And backtests? Finding rules that worked in the past is easy—finding ones that weren’t just overfitting noise is where the bodies are buried.

“Improving” the system is even harder:

This is kinda interesting because “you should iteratively improve your trading system” is excellent advice for a professional operating in a professional environment with the discipline and tooling that prevents them from overfitting (indeed it’s basically mandatory — if you…

— cephalopod (@macrocephalopod) March 23, 2025

At the first drawdown, doubt creeps in. Has the world changed? Look at VIX. For decades, it wore the crown as the “fear index.” But now? With most of the fireworks happening in 0DTE options, is VIX still the right gauge? Are you equipped to answer that?

If you’ve followed the logic so far, you’d expect me to steer clear of regime-based strategies. And yet, I can’t stop looking.

Here is how the paper authors describe their methodology:

“Our paper proposes a systematic approach to regime selection. The user of our method needs to specify a set of economic variables. In our empirical example, we consider a constellation of seven variables. We transform these variables to look at annual changes and compute a z-score. Variable by variable, we identify times in the past that are similar. Our measure of similarity is the squared distance of today’s value to each historical observation. For example, if the z-score today is 2.5, we look at historically similar times where the z-score is close to 2.5. If the value on a particular historical date was exactly 2.5, the squared distance would be zero. We then look at every historical date and aggregate the distances at each date across our seven variables. We refer to the aggregated similarity score as the global score. Those historical dates with the smallest aggregate distances (meaning that they are most like today) are our definition of similar regimes. Once we have established similar dates in the past for a particular asset, we look at subsequent returns. So, if we observe subsequent historical returns that are very positive, this suggests that the ex-ante returns for that asset today are also positive. Importantly, once the historical regimes are established, we can apply this method to any asset.“

The idea is valid and rather simple to understand and implement. The authors also realise the many pitfalls in it:

“Choosing these variables – as with more parametric approaches – induces a type of look-ahead bias. Today, we know which variables have been important in the past. We can partially mitigate this issue by choosing variables that were important before our sample begins. Second, there is a choice as to how to represent the variable; for example, should we detrend it? If looking at rate of change, over what horizon? Third, we need to take a stand on the degree of similarity. For instance, is a z-score of 2.3 similar enough to 2.5? Fourth, we need to decide on the length of the observation period after the similar historical regime. Finally, there is a decision on how to weight the economic variables – we choose equal weights but that is a choice.“

“For our empirical illustration, we use the seven economic state variables detailed in Exhibit 2: the
S&P500 index level, 10-year bond yield minus 3-month treasury bill yield (slope of the US yield
curve), WTI crude oil price, copper price, US 3-month treasury bill yield, VIX (pre-pended with
realized volatility before 1990), and the US stock-bond correlation. For each variable, we take a
12-month change and then normalize it by computing the z-score over a rolling 10 years, capped
to be within -3 and 3.“

Feels like a typical machine-learning problem? You have a list of possible variables, attributes of those variables and ranking methods. But as I wrote above, the onus is still on us to judge when we found a reasonable set of variables vs an overfitted concoction that would not work from day 1. Can we apply the Lindy effect here?

“Although we refer to these as economic state variables, they are all financial variables.“

I do not know why they chose to call them economic variables and then add the above line but here we are…

It is true that financial variables co-move with economic ones but the relationship is never 1-to-1. Having both sets of variables equally available, I was puzzled about the choice…until I remembered that my preferred model, Verdad’s Best Macro Indicator, is also a financial variable.

And really, this isn’t an econ paper. The goal isn’t to explain the world—it’s to make money. If financial variables get you to the right call faster, that’s what matters.

“The low average absolute cross-correlations suggest that the seven economic variables chosen
contain independent information.“

Once the model identifies the regime, the authors proceed as follows:

“we do apply our method to six popular long-short equity factors. We look at historical factor returns and go long a factor if the observed historical return subsequent to observing the regime at the investment date is positive, and short otherwise.“

Why do they look at equity factors and not asset classes? ¯\_(ツ)_/¯

“We illustrate the efficacy of our methodology on six long-short stock factors: using the FamaFrench five research factors (Market, Size, Value, Profitability and Investment) plus the 12-month
Momentum factor. Essentially, we are testing whether our systematic regime classification could
be a useful tool for factor timing.“

“The one thing that we know is that factor timing is difficult as emphasized by Asness (2016) and
Asness et al. (2017). Many of the previous approaches are highly parametric and open the risk of
overfitting. Our approach is much more non-parametric.“

The Holy Grail. Factors work—until they don’t. And sometimes, they don’t for a painfully long time. If you can time them, you get the double win: extra profits and job security. Nothing keeps you employed like avoiding those stretches where your strategy lags the market.

Here is where the paper gets…weird? I get it—they’re just showcasing a new way to identify regimes. This is a test run. But the long-only model already has the best Sharpe and returns… so, what are we doing here?

The authors argue that going long the best quintile and short the worst gives you a high Sharpe and low correlation to the long-only portfolio. But the long-only portfolio is already long beta plus a long/short factor tilt. Isn’t that already the best of both worlds? End of the day, low correlation only matters because you want something to pair with stock beta… right?

The strategy doesn’t even pass the eye test—the first 20 years were great, the last 20… not so much. Same story as a lot of long-only factor strategies, and we all know how that played out for money managers. If this thing delivered a smoother equity curve, I’d be all in. But like this…?

Then again, the authors work for one of the best hedge funds out there. So who am I to say anything?

What I am reading:

Follow me on Bluesky @nprotasoni.bsky.social