The ideas behind the new Ballast: where markets can change character, how we measure the warning signs, and an honest account of which parts we have actually tested.
Markets don't move in straight lines. For long stretches they are calm and forgiving; then, sometimes quickly, they change character: falls get faster, everything starts moving together, and the diversification you counted on shrinks. Ballast is built around that observation. It does not try to predict when the next storm arrives. It does two other things: it watches for the conditions that have historically come before rough patches, and it shows you how your own mix would have coped in past storms, so you can decide how much risk you are comfortable carrying before one arrives.
The usual portfolio toolkit assumes a smooth world: risk is one number, correlations are stable, and bad days are rare and independent. In real crises, bad days cluster, correlations jump toward one, and a mix that looked well diversified falls together. A single "risk score" cannot tell you whether the market is getting more fragile or whether your particular holdings are exposed to it. Those are two different questions, so Ballast keeps them as two separate gauges.
The simplest system that shows how order turns into chaos is the logistic map, a one-line population model. Turn one dial up and its steady state splits into two, then four, then eight values, ever faster, until behaviour becomes effectively unpredictable. In the 1970s Mitchell Feigenbaum found that the spacing of those splits follows a universal ratio, about 4.669, that appears in very different systems.
How Ballast uses it: as a way of describing a score, not as a law of markets. Ballast's early-warning score is divided into named bands (Stable, Oscillating, Reactive, Chaotic) with a small buffer so the label does not flicker between bands from day to day. Nobody has shown that financial markets literally follow period-doubling; the band names are a plain-language scale.
René Thom's catastrophe theory (1970s, popularised by Christopher Zeeman, who applied it to stock market crashes) describes systems with a "fold": over one range of conditions there are two stable states, and the system can stay in one until a small change tips it suddenly into the other, and it does not return along the same path. That is the picture behind "slow build-up, sudden break".
How Ballast uses it: as an illustration in the Learn layer. It is a qualitative model; we do not estimate its parameters from market data and we do not claim it forecasts crashes.
In 1987 Per Bak, Chao Tang and Kurt Wiesenfeld showed that a pile of sand fed one grain at a time organises itself into a state where the next grain may do nothing or trigger an avalanche of any size. It is a vivid picture of why small events sometimes have enormous effects, and why quiet periods can hide a build-up.
But "fragility" cannot be read straight off a theory. In Ballast it is built only from things we can measure from prices: how strongly shocks cluster (Hawkes intensity), how much correlation rises when markets fall (tail dependence and dynamic correlation), how concentrated the real risk is (effective number of independent bets), and how surprised a market-state model is by recent moves. We do not measure liquidity, leverage or crowding, and Ballast does not claim to. The fragility meter is marked experimental.
Systems near a tipping point recover more slowly from small knocks, so their fluctuations become larger and more persistent just before the shift (Scheffer and colleagues, Nature, 2009). In markets the practical signals are rising variance, rising lag-one autocorrelation, and "flickering" between two states. The Portfolio Research Console computes exactly these.
How the stability signals fared when tested: we ran a pre-registered test on the S&P 500 and the ASX 200 and found they add nothing once momentum is already included, so they are shown as experimental descriptions, not scored. We also tested one extra idea, "quiet but fragile" (a calm basket whose stability signals look stretched). It did not hold in both periods and did not mark a high-risk state, so it was not added.
What did work: the Ballast weather score, built from eight market conditions. Since 1990 on the S&P 500, a fall of 10% or more within about six months followed roughly 4% of Calm days, 21% of Stirring, 34% of Oscillating, 38% of Reactive and 59% of Stormy days (22% of all days). It separates falls modestly on the S&P 500 and only weakly on the ASX 200, and it says nothing about direction. Separately, the Research Console's learned risk read for a basket beat the plain base rate out-of-sample (Brier skill of roughly +7%, lower end of a 90% range about +4%). That is the Console's own model, not the weather score.
Underneath the regime layer sit methods already explained elsewhere in this hub: Ledoit–Wolf shrinkage and random-matrix denoising to clean the correlation matrix, robust SOCP so the allocation is not hostage to estimation error, hidden-Markov regimes, CVaR for the size of tail losses, and copulas for crash-time co-movement. These are standard techniques in quantitative finance, not Ballast inventions.
Instead of a promise, Ballast shows a rehearsal. It replays a mix through past storms using only the information available at the time (walk-forward, no hindsight). Below is the worst fall inside each window on the ETF universe back to 2013, for the three risk tiers (the tier sets how strongly the protective overlay is applied).
One important caveat: these figures are measured at month-ends, so they understate how far a portfolio fell between peak and trough. The Australian market's COVID fall was about 21% on this basis and about 36% measured day to day. Read the table as a comparison between settings, not as the size of a fall. The Ballast app's own replay uses daily prices.
| Storm | Defensive | Balanced | Aggressive | Overlay off | Fixed mix | Market |
|---|---|---|---|---|---|---|
| COVID (Feb–Mar 2020) | −6.7% | −7.8% | −8.5% | −9.7% | −10.8% | −20.7% |
| 2022 (whole year) | −8.5% | −10.5% | −12.4% | market −12.0% | ||
On longer index-based proxies reaching back to 2006 the tiers helped in the sudden 2020 crash (market −21.2%, tiers −10.8% to −13.5%) but did not help in 2022, a slow grind driven by rising interest rates (market −13.7%, tiers −15.4% to −21.8%). Over the whole period each step toward the defensive tier gave up roughly 0.5 to 1.5 percentage points a year of return in exchange for a shallower worst fall.
A fifth "periodic windows" regime failed its test. The idea was that after chaos, a short stretch of structured opportunity opens. We wrote the rule down in advance and tested it on the S&P 500 and the ASX 200: forward six-month returns after such windows were no better than average (S&P 500 +3.6% versus +4.7% for all months; ASX 200 +2.6% versus +2.7%). So it stays a story label for recovery, and Ballast does not change an allocation because of it.
The signals describe conditions; they do not reliably time them. The protective overlay is better thought of as insurance, with a cost, than as a source of extra return. The backtests rest on a small number of real storms, and past behaviour does not guarantee future results. Nothing here predicts crashes, and none of it is a recommendation to buy or sell anything.
The logistic map became famous through the biologist Robert May's 1976 Nature article "Simple mathematical models with very complicated dynamics", partly a plea that everyone who works with equations should play with this one. Feigenbaum, a physicist at Los Alamos, then found its universal constant.
Zeeman's attempts to model stock market crashes with the cusp catastrophe were later criticised as hard to test, a fair warning about how far such pictures can be pushed, and the reason Ballast labels them as illustrations. Bak's sandpile, by contrast, started as a computer model and became one of the most cited ideas in complexity science.
General education only. Nothing here is personal financial advice or takes into account your objectives, financial situation or needs. Past performance is not a guarantee of future results.