ORCA Is Live

Let me start with the part nobody likes to say out loud.

Almost every risk model deployed by almost every serious institution in the world is, at its core, a glorified rearview mirror. It tells you the road got bumpy after you have already hit the pothole. We dress this up in Greek letters and conditional variances and call it sophistication, but the brutal truth is that volatility-based risk systems detect crises the same way a smoke alarm detects fire: too late to save the furniture.

Today, with my brother Fedor at UTS, I'm releasing ORCA — the Online Regime Correlation Analyzer. The paper is on arXiv, the live dashboard is up, and I want to use this post to say a few things I could not say in a peer-reviewed manuscript.

The premise is almost embarrassingly simple

Stop watching how loud prices are screaming. Start watching how synchronized they have become.

That is it. That is the whole idea. Markets in good health look like a sprawling, messy, somewhat chaotic web — bonds doing bond things, gold doing gold things, tech ignoring utilities, emerging markets marching to their own drum. Markets approaching a crisis look like a marching band. Everything tightens. Diversification, that sacred concept your financial advisor has been selling you since 1985, quietly evaporates before anyone announces it has evaporated.

The eigenvalues of the cross-asset correlation matrix know this. The clustering coefficient of the dependence graph knows this. The spectral gap knows this. None of these things show up on Bloomberg's front page. We extract 127 of them, glue them to 79 traditional indicators, hand the resulting 206-dimensional vector to a Random Forest, and ask it a single question: what is the shape of risk right now?

The numbers, briefly, because someone always asks

On a strict eight-fold walk-forward backtest covering 15 years and including COVID and the 2022 rate-hike drawdown:

BCD-AUC of 0.741 — first place against every baseline we tested, including HAR-RV, turbulence indices, and traditional-feature random forests. Spectral features add +10.3 points of AUC for crash detection. Not 1 point. Not 3. Ten. Strategy Sharpe of 1.13. CAGR of 15.6%. Maximum drawdown of −7.5%. Buy-and-hold over the same window: Sharpe 0.09, CAGR 3.7%, drawdown −33.7%.

I want to be honest about what this means and what it does not. It does not mean we have solved markets. Nobody has solved markets. Anyone who tells you they have is either lying or about to blow up. What it does mean is that there is a class of information — the topology of how things are connected — that the entire industrial complex of risk management has been systematically under-weighting for thirty years, and that information is recoverable in real time with mostly-public tools.

Why I'm putting it on the open internet

Because I am tired of the genre.

Look at how this kind of work usually surfaces. A hedge fund builds something. They keep it private. Eventually a glossy article appears in a trade publication describing how this fund "uses proprietary network analysis" to "navigate complex market regimes." A pension fund pays seven figures for access. The methodology is never disclosed. The result is never independently verified. The community never gets to challenge it, improve it, or prove it wrong.

That cycle is bad for science and bad for everyone who is not already inside the velvet rope. So we wrote down what we did, what features we used, how we validated, what worked, what didn't, and what the failure modes are. The dashboard at orca.boriskriuk-powered.com refreshes every twelve hours on live data. You can stare at the same correlation graph I stare at. You can disagree with our regime mapping. You can fork the methodology and beat us. Please do.

The thing the SHAP plots told me that I did not expect

If you only read one figure in the paper, make it the SHAP analysis. Here is what it shows, and here is what it shocked me into believing:

Rallies and crashes are not symmetric. They are not the same coin flipped two ways. Rallies are mean-reversion events — they show up in drawdown indicators, in price-to-moving-average ratios, in the boring old technical features that retail traders have been drawing on charts since the 1970s. Traditional features dominate the rally model.

Crashes are something else entirely. The top three features in the crash model are all graph-topological: the clustering coefficient at correlation threshold 0.5, the edge density at the same threshold, and the percentile rank of the dominant eigenvalue. None of them look at price. They look at structure. They are asking: how tightly is the market gripping itself right now, relative to how tightly it has gripped itself over the last year?

The implication is uncomfortable for a lot of practitioners. It says crashes are not, fundamentally, a price phenomenon. They are a connectivity phenomenon that subsequently expresses itself in price. By the time it reaches price, it is already a memory.

What I think this changes, and what I do not

I am going to resist the temptation to tell you ORCA is a revolution, because I have read enough academic papers that promised revolutions to last me a career. Most papers do not survive contact with live capital.

Here is what I will say. If the structural-precursor hypothesis holds — and the ablation study is, to my eye, the strongest internal evidence we have that it does — then a few things should happen in the next several years. Pension funds should start treating correlation topology as a first-class risk metric, not a footnote in a quarterly report. Regulators monitoring systemic stress should be looking at spectral collapse, not just at VIX prints. And the comforting story that "I'm diversified, I'm fine" should come with an asterisk: you are diversified only as long as the network stays loose.

The dirty secret of 2008, of March 2020, of every textbook crisis, is that diversification stopped working at the exact moment people needed it to work. ORCA is a tool for watching that failure happen in real time, before the price action confirms it.

Where I think we are wrong, or might be

A few honest caveats, because I refuse to write the version of this post that pretends we nailed it.

We tested on US markets only. Whether spectral crisis precursors generalize to Europe, Asia, or genuinely fragmented emerging markets is an open question. The Random Forest is interpretable and robust, but a more aggressive learner might extract more juice from the same features — at the cost of explainability, which I am not willing to give up yet. And every backtest in the history of finance has flattered itself. Live deployment is the only honest test, and we are running that test now, in public, with timestamped predictions you can check against reality as it unfolds.

A closing thought

The deeper lesson of this project, for me, has nothing to do with finance. It is about what it means to measure a complex system. We have spent a generation collapsing markets, ecosystems, brains, and economies into single numbers — single GDP figures, single volatility readings, single happiness indices — and then acting surprised when those numbers fail to warn us of the things that matter. The information was never in the number. It was in the pattern of relationships the number was summarizing away.

If ORCA does anything, I hope it nudges the field toward listening to the shape of the web rather than the scream of any single thread. The web tightens long before it tears. We just have to be looking.

The paper is here. The dashboard is here. Come break it.