University of Oxford Claims You Can Predict Market Crashes a Week Ahead
A new research paper out of the University of Oxford claims that using a hybrid machine-learning model, you can forecast market “crashes” a week in advance.
The results:
40% annual returns
–18% max drawdown
Beta 0.51 (relatively low market exposure)
These numbers are good on paper, but that’s not what caught my attention the most.
What I found interesting is where the author’s crash signals actually came from.
Author Claims Stocks Don’t Lead, They Follow
According to this research, SPY is the last place you see declines before market crashes.
The paper finds that the earliest warning signs of a selloff show up in other asset classes first, long before equities react.
Here are the author’s top signals:
Oil markets
Extreme skew or kurtosis in oil tends to appear before stock weakness suggesting macro stress arrives through commodities first.
Currencies
FX markets react instantly to global risk. Sudden moves in USD, EUR, or JPY often precede equity turbulence.
Bond yields & short-term rates
Rapid changes in interest rates show up ahead of SPY drawdowns. The credit market prices stress faster than equities do.
Small caps (IWM) crack before big caps
Small-cap equities which are more liquidity-sensitive and more economically exposed tend to panic first. Large caps lag.
Where CI Volatility Disagrees
The author defines a “crash week” as:
A 1% decline in SPY over a 5-day window.
To us, that is not a crash.
That is normal, everyday volatility.
SPY moves ±1% all the time.
That definition inflates the number of “crashes”.
What We Will Be Testing
Despite not agreeing with the definition of a “crash”, the cross-asset theory behind it is fascinating.
So here’s what we’re doing next:
First, we’ll test each of the author’s claims independently (oil only, FX only, rates only, IWM only)
Then, Pair combos (oil + IWM), (FX + rates), etc…
Then, Triplet combos (oil + IWM + FX), (oil+ rates + IWM), etc…
Finally, all four together.
Next, for every combination, we’ll look at different weightings (for example 70% IWM / 30% oil or 60% FX / 40% IWM, or all equal-weight).
Finally, we’ll have to test each scenario across different VIX regimes.
Only the versions that show statistically significant predictive power for real drawdowns (≥5%, not 1% everyday moves) will survive.