How TRKR's simulation engine works

What it produces

The engine generates hundreds of alternative "futures" for real markets — daily prices for 20 assets, simulated jointly with the macroeconomic environment that drives them (the policy rate and Treasury yields, credit conditions, equity and bond-market volatility, the dollar). Each future starts from today's actual levels of those variables and evolves forward day by day for up to ten years. The market regimes and the macro environment are learned from more than three decades of history, from 1990 on, which spans several interest rate cycles. Each asset's behaviour is learned from the two decades its funds have traded.

It is important to be clear about what this is not: it is not a forecast. No single simulated future is a prediction. The value is in the ensemble — the full range of what could plausibly happen — which is what you need for stress testing, retirement sequencing, drawdown analysis, and judging whether a portfolio is robust rather than merely optimized to the one history that happened to occur.

The core idea: markets have weather, and weather has seasons

Most simple simulators treat returns as coming from one fixed distribution — same volatility, same correlations, forever. Real markets don't work that way. Volatility clusters. Diversification that works in calm markets famously evaporates in crises, exactly when it's needed most. Bonds sometimes hedge equities and sometimes don't, depending on what's driving the stress.

The engine is built around this reality using a three-layer cascade:

1. The market regime comes first. The engine identifies three broad regimes in history — a calm bull state, a choppy sideways state, and a stressed bear state — from the indicators that mark stress: volatility, credit spreads, and the shape of the yield curve. Regimes persist realistically (a crisis lasts weeks or months, not a day) and switch at the frequencies observed since 1990. Each future draws its own sequence of regimes, so the ensemble covers calm and stressed decades in their historical proportions.

2. The macro environment follows the regime. Interest rates across the curve, credit spreads, equity and bond volatility gauges, and the dollar are simulated as one coherent, interlinked system, with each regime carrying its own typical levels and behaviour — so a simulated rate shock arrives with the credit and volatility behavior that historically accompanies rate shocks, never in isolation. Because this layer learns from 1990 on, the level interest rates gravitate toward reflects several cycles, not just the near-zero-rate years after 2008.

3. Assets respond to both.Each asset's returns are driven by its sensitivity to the macro moves — and, critically, those sensitivities change with the regime, because that is what the data shows: a risk asset's response to a volatility spike, or a bond fund's effective rate sensitivity, is not the same in a bull market as in a panic. On top of this macro-driven component, the engine layers each asset's own dynamics and the co-movement between assets — again regime by regime — including the fat tails and asymmetry (crashes are sharper than rallies) that normal-distribution models miss.

The practical consequence: in the simulated bear states, correlations rise, volatility spikes, and losses have realistic depth and speed — so a portfolio that looks diversified "on average" gets tested against precisely the environments where diversification tends to fail.

Discipline: calibrated to evidence, guarded against overfitting

Every quantity in the model is estimated from the real joint history of the assets and the macro series — nothing is hand-set to produce a desired answer. Where the data is thin (rare regimes, short samples), the estimation deliberately leans conservative rather than extrapolating aggressively. The specific estimators, stabilization techniques, and parameter choices are proprietary; they are the product of extensive research iteration.

Quality control before anything is published

Realism is not taken on faith internally. Before a simulated dataset reaches the platform, it is scored against real history using the same yardsticks on both sides: per-variable drift, volatility, and worst-drawdown behavior; the full cross-variable correlation structure; whether asset sensitivities rotate across regimes the way they do in history; and checks that the simulator doesn't manufacture spurious predictability — patterns that would let a strategy "see" tomorrow more clearly than real markets allow, flattering any backtest run on it. Every published dataset carries that scorecard on the datasets page, so where a dataset drifts from history, the drift is shown rather than hidden.

Honest limitations

The engine reproduces the kindsof environments present in its training history, including its crises. It cannot invent a genuinely unprecedented structural break with no historical analogue. Its stress states are as severe as history's — a feature for realism, but judgment still applies for truly out-of-sample scenarios. And again: it describes the range of outcomes, not which outcome will occur.

Two limitations are specific and measured. First, the assets' behaviour is learned from 2006 on, an era in which high-grade bonds cushioned equity losses in almost every year. Years in which stocks and bonds fall together, as in 2022, do occur in the simulated futures — roughly one year in twenty — but less often than recent markets would suggest. Second, the engine does not simulate inflation or GDP growth: neither has a daily print, and in testing neither improved the model's fidelity out of sample. Their influence reaches the futures through rates, credit conditions and volatility.