Can Historical Optimisation Deliver Robust Financial Models?

Financial markets generate enormous amounts of data, yet the number of genuinely distinct market environments available for evaluating financial models is surprisingly limited. Major liquidity crises, sustained inflationary tightening cycles and other periods of severe disruption are relatively rare.

This creates a tension at the heart of quantitative finance. We routinely optimise strategies, portfolios and models against historical observations, but how much confidence should we place in results obtained from a limited range of market conditions?

When optimisation meets limited history

Optimisation is central to quantitative finance. Trading strategies, portfolio allocations, risk parameters and model hyperparameters are routinely selected or adjusted using historical data.

There is nothing inherently wrong with this. The difficulty arises when extensive optimisation allows a model to adapt not only to persistent relationships in the data, but also to the particular circumstances and chance fluctuations of the observed historical period.

A strategy may perform exceptionally well in a backtest because it has adapted to characteristics of the observed historical period. A risk model may appear reliable because certain forms of market stress were rarely encountered. A portfolio may appear well diversified because the historical sample contains relatively few episodes of widespread correlation increases.

In each case, strong historical performance may coexist with substantial vulnerability to conditions that were poorly represented in the data.

This does not mean that optimisation is necessarily misguided. It means that historical performance and robustness are different properties, and success at one does not establish the other.

Beyond the observed market path

History is the only realised sequence of market events directly available to us. But it is not the only sequence that could plausibly have occurred.

The future will not simply replay the past, yet neither will it be unconstrained. Financial markets exhibit persistent statistical and temporal characteristics, including heavy-tailed returns, volatility clustering and dependencies between assets. These characteristics provide useful information about the kinds of behaviour that alternative market trajectories might exhibit.

Generative modelling offers a way to explore such alternatives. Rather than evaluating a model or strategy solely against the observed history, we can examine its behaviour across an ensemble of synthetic trajectories designed to reproduce relevant characteristics of financial markets.

The objective is not to predict which trajectory will occur. It is to investigate how sensitive a model’s performance may be to different plausible sequences of market events.

What can synthetic scenarios establish?

Testing across synthetic histories may reveal sensitivities that are not apparent in a conventional historical backtest, particularly when those histories contain different sequences of volatility shocks, extreme returns and changing cross-asset dependencies.

However, synthetic data do not create genuinely independent evidence about market conditions that have never been observed. Generated scenarios inherit limitations from the historical data, modelling assumptions and structural constraints used to produce them.

Nor does reproducing selected statistical characteristics guarantee that every generated trajectory is plausible. The credibility of synthetic scenarios requires careful examination of their distributional properties, temporal behaviour and dependencies, together with an understanding of what the model does not capture.

Scenario analysis is therefore a complement to historical evaluation, not a substitute for it.

From historical performance to robustness

Optimisation can tell us which models or strategies perform best against the evidence available. It cannot, by itself, establish how robust those results will be when market conditions differ from those observed.

Exploring alternative plausible market histories offers a way to investigate that dependence. The objective is not to replace historical optimisation, but to understand its limitations and assess how sensitive its conclusions may be to the particular history on which they were based.