Financial time series are non-stationary in a deep sense. The statistical properties of asset returns — their mean, variance, autocorrelation structure, and cross-asset dependencies — evolve over time in response to macroeconomic shifts, regulatory changes, liquidity conditions, and market microstructure dynamics. A model calibrated on data from one market regime may produce predictions that are not merely inaccurate but structurally invalid when applied to data generated under a different regime.
This represents a fundamental challenge for machine learning approaches that assume stationarity, or that treat regime heterogeneity as noise to be averaged over rather than signal to be identified. Black-box models optimised for in-sample fit are particularly vulnerable: they can learn spurious correlations specific to a particular regime and fail completely when that regime ends.
Before any predictive model can be responsibly deployed in a non-stationary environment, the concept of regime must be made operationally precise. A regime, in this context, is a period during which the data-generating process is approximately stationary — a local window within which a fixed model provides a valid description of the dynamics.
Identifying regime boundaries is therefore a prerequisite for principled model construction. When the regime is known or detectable, the modeller can:
The requirements of quantitative finance impose constraints on model form that black-box methods routinely fail to satisfy.
Under frameworks such as SR&ED, MiFID II, and proposed algorithmic accountability regulations, investment firms are required to explain the decisions produced by algorithmic systems. A symbolic model — an explicit mathematical expression — satisfies this requirement directly. A gradient-boosted ensemble does not.
A symbolic expression encodes a structural relationship. If that relationship is grounded in fundamental economic mechanics rather than historical correlation, it may remain valid across regime transitions in a way that a statistically fitted function cannot.
Risk managers require the ability to stress-test models analytically — to ask what happens to model output as specific input variables are perturbed. This requires differentiation of the model function, which is trivially available for symbolic expressions and non-trivially available for neural networks.
The application of symbolic regression to financial data requires adaptation of the standard methodology to account for temporal dependence and non-stationarity. Key modifications include:
Rather than applying SR to the full historical dataset, we apply it to rolling windows of data, producing a sequence of candidate expressions that characterise the dynamics within each temporal segment. This allows the symbolic structure of the relationship to evolve over time while retaining interpretability within each window.
When the symbolic expression that best describes the data changes abruptly between successive windows, this signals a regime transition. The nature of the structural change — which terms appear or disappear, which coefficients shift — provides direct information about the economic mechanism underlying the transition.
Raw price and return series are replaced by features constructed to be stationary by construction — ratios, spreads, rank-transformed quantities — before symbolic search is applied. This mitigates spurious discovery of relationships that are artifacts of common trends rather than structural dependencies.
In a study of equity volatility dynamics, applying symbolic regression within distinct volatility regimes identified by a Hidden Markov Model yielded structurally different expressions for the two regimes:
The qualitative difference between expressions is immediately interpretable: high-volatility regimes are characterised by stronger return-shock amplification and a nonlinear self-reinforcing term absent in low-volatility periods. This kind of structural insight is inaccessible in black-box volatility models.
Interpretable AI is not an academic luxury in quantitative finance; it is a practical necessity. The combination of regulatory requirements, model risk management standards, and the fundamental non-stationarity of financial data creates a setting where symbolic, transparent models offer genuine operational advantages over opaque statistical estimators. Symbolic regression, applied within a rigorous regime-aware framework, provides a methodology for extracting structural relationships from financial time series that are both operationally deployable and scientifically credible.