Luigi Piva — ALGOSWORKSAI LTD
This is the complete, honest account of a week of testing on a simple question: can you use a statistical measure of market stress to improve an equity portfolio, exiting into turbulence and re-entering into calm? The short answer is “yes, but.” The long answer — with the dead ends, the mistakes, and the corrections — is more useful than the short one, and I’m telling it in full because that’s how serious validation is done.
The starting point
A SPY buy & hold is a tough opponent. Over long periods it generates a Sharpe around 0.9 with a high, positive risk premium. Any strategy that reduces exposure starts at a disadvantage: every day out of the market is potential return lost. To beat buy & hold on a risk-adjusted basis, the moments when you exit have to be genuinely the right ones.
The hypothesis: turbulence indices (Turbulence Index, Absorption Ratio) identify those moments.
Stage 1: the single-asset signal isn’t enough
My first mistake, an instructive one. I tried computing Turbulence on SPY alone. But Turbulence is a multivariate measure: on a single asset it degenerates into the squared z-score of the return, and the Absorption Ratio isn’t even defined (one asset = one eigenvector). The method’s informational richness lies in the correlation structure between assets — which simply doesn’t exist for a single security.
Lesson: the signal must be computed on a basket (the seven asset classes), even if you then invest in a single instrument.
Stage 2: the wrong calibration
Correct multivariate signal, computed on the seven assets. First scaling calibration: exposure = 1 − turbulence percentile. Continuous, proportional.
Result: Sharpe 0.68 against 0.89 for buy & hold. Worse than doing nothing. Average exposure collapsed to 30–40%.
I was exiting too early and too often. In a rising market, holding 60–70% of capital in cash “for prudence” is devastating. The drawdown reduction didn’t offset the return lost.
Here many people stop and write “the technique doesn’t work.” That would have been a mistake.
Stage 3: a bug worth telling
Moving to the full multivariate test, I got a table where all strategies — at any exposure level — gave CAGR, Sharpe, and drawdown identical to buy & hold. Immediate suspicion: a portfolio exposed at 30% cannot return like one at 100%.
The bug: in the performance function I computed net return without multiplying by exposure. One missing line — gross = exposure * returns — and every strategy pocketed SPY’s full returns regardless of how invested it was.
I tell this not for self-flagellation but because the way I found it is the point: by not trusting a too-good result. The sanity check “are these numbers physically possible?” is worth more than a thousand lines of elegant code. A backtest you don’t challenge is a backtest that lies to you.
Stage 4: the right calibration
With the bug fixed, I used the paper’s actual calibration: quintile scaling. Fully invested up to the 60th percentile of turbulence; reduction only in the tail (75%, 50%, then out in the extreme).
Finally: Sharpe above buy & hold, drawdown nearly halved, average exposure at 83%. The exact same technique that failed in Stage 2 works here. The only difference was where I placed the exit threshold.
In the exact period of the reference article (from 2013 due to data constraints), the Sharpe rose to 1.00 against 0.92 for buy & hold.
Stage 5: the Absorption Ratio and the limit of fit
I also tested the Absorption Ratio with its leading signal (standardized shift), including 2008. It worked as promised on timing: during the crisis exposure dropped early, drawdown went from −55% to −40%. But the Sharpe stayed below buy & hold.
The reason is structural: long-only equity has a risk premium that dominates. A signal that protects you from tails gives its best on instruments with asymmetric negative tails — FX carry, for instance, as in Kritzman’s original application. On SPY the protective value is there but not enough to overcome the cost of staying out.
What I learned, in order of importance
- Calibration matters more than the technique. Same signal, opposite outcomes depending on the threshold. Most of the value (or damage) lives there.
- Instrument-signal fit is decisive. “Does it work?” is the wrong question. “Does it work for this instrument and this objective?” is the right one. Drawdown reduction on equity: yes. Alpha on equity: marginal. Overlay on tail strategies: promising.
- Don’t trust the pretty results. The bug was caught by physical common sense, not by the debugger. Every backtest must be interrogated with “is this possible?”.
- In-sample stays in-sample. All this work chose parameters by looking at the test data. The gap between my Sharpe of 0.95–1.00 and the paper’s 2.20 remains partly unexplained, and the real validation — walk-forward out-of-sample — is still to be done. These numbers are exploratory.
Conclusion
Regime-detection techniques based on turbulence and factor compression are real, useful tools for risk management. They improve a SPY buy & hold — modestly, but verifiably — if calibrated to exit only in the tail. They’re not alpha machines, and whoever sells them as such is hiding the calibration, the period, or both.
The value, for an allocator, isn’t a magic number. It’s the ability to know which regime you’re in and act accordingly, with full awareness of where the method helps and where it doesn’t. This — validating with honesty, not selling with enthusiasm — is the craft.
ALGOSWORKSAI LTD develops and validates systematic strategies for institutional allocators. The results presented are exploratory and in-sample; they constitute neither a track record nor investment advice.
Get the monthly Market Regime Note
Regimes, volatility and correlations across major futures markets — with the code behind the charts. Free.
Subscribe →