Fraction of Variance Unexplained

Measuring Model Limits in Financial Markets

Quantitative models in financial markets are built to explain variation. Whether forecasting returns, estimating risk, or identifying structural relationships, these models attempt to impose order on inherently noisy data. Their performance is typically summarised through measures of fit, i.e. how much of the observed variation they successfully capture.

Yet an equally important perspective lies beyond what is explained:

the proportion of variation that remains outside the model.

This is formalised through the concept of the Fraction of Variance Unexplained (FVU). Rather than emphasising explanatory power, FVU directs attention to model limitations. In financial markets; where uncertainty, noise, and structural change are persistent; this perspective is not merely complementary; it is essential.

Defining the Fraction of Variance Unexplained

The Fraction of Variance Unexplained measures the share of total variability in a dataset that is not captured by a model. It is typically expressed as the ratio of the variance of residuals (model errors) to the total variance of the observed data.

In practical terms, FVU answers a direct question:

How much of the data lies beyond the model’s explanatory reach?

An FVU close to zero indicates that a model captures most of the variation. A higher FVU suggests that a significant portion of the system remains unaccounted for.

This framing shifts the analytical focus. It moves from evaluating success to acknowledging limitation.

Interpretation in Financial Markets

In many scientific domains, a low FVU is considered unambiguously desirable. In financial markets, the interpretation is more nuanced.

Markets are shaped by behavioural dynamics, strategic interaction, exogenous shocks, and continuous structural evolution. As a result, a meaningful proportion of variance is inherently difficult, if not impossible, to explain.

A low FVU may reflect strong explanatory power. Equally, it may indicate overfitting to historical data. Conversely, a higher FVU may signal genuine uncertainty, noise within the system, or limitations in the modelling framework.

This ambiguity is not a flaw in the metric. It is a reflection of the environment in which it is applied.

Noise, Signal, and the Structure of Residuals

Residuals are often treated as random noise—irrelevant deviations from the model’s predictions. In financial markets, this assumption is frequently incomplete.

Unexplained variance can arise from omitted variables, changing relationships, non-linear dynamics, and behavioural effects. What appears as noise may, in some cases, represent latent structure that has not yet been captured.

This distinction is critical. Not all unexplained variance is meaningless. Some of it may contain information about emerging patterns, regime-dependent behaviour, or structural shifts.

FVU therefore does more than measure error. It highlights the boundary between what is understood and what remains to be explored.

Model Complexity and the Illusion of Fit

Model complexity has a direct influence on FVU. Simpler models often leave more variance unexplained, resulting in higher FVU. Increasing complexity can reduce FVU by capturing additional relationships within the data.

However, this reduction is not inherently beneficial.

More complex models introduce risks: overfitting to historical data, reduced interpretability, and instability across changing market regimes. In dynamic environments, models that appear highly accurate in-sample may fail when conditions shift.

FVU thus reflects a fundamental trade-off. Lower unexplained variance may indicate a better fit, but not necessarily better generalisation. In financial markets, where adaptability is critical, this distinction matters.

FVU as a Reflection of Uncertainty

FVU is closely tied to the concept of uncertainty. In systems with stable, well-defined relationships, unexplained variance can be minimised. Financial markets, however, are characterised by relationships that are dynamic, context-dependent, and influenced by human behaviour.

This leads to persistent residual variation.

In this sense, FVU can be interpreted as a proxy for model limitation, structural uncertainty, and the irreducible complexity of the system. It quantifies not only what the model fails to capture, but also what may not be fully capturable.

Time Variation and Regime Sensitivity

Unexplained variance is not static. It evolves over time and across market regimes.

During stable periods, relationships may strengthen, and FVU may decline as models align more closely with observed data. During periods of disruption, correlations can break down, volatility rises, and FVU typically increases.

This dynamic behaviour provides valuable insight. Rising unexplained variance may signal regime shifts, weakening relationships, or the breakdown of existing modelling assumptions.

In this way, FVU functions not only as a diagnostic of model performance, but also as an indicator of changing market conditions.

Implications for Quantitative Systems

Within quantitative frameworks, FVU serves as a diagnostic tool that extends beyond traditional performance metrics. It supports the evaluation of model robustness, highlights structural limitations, and provides context for interpreting predictive signals.

Its value, however, lies not in minimisation alone, but in interpretation.

Understanding the sources of unexplained variance, assessing its stability over time, and recognising its implications for decision-making are all critical. Without this context, FVU risks being reduced to a single number rather than a meaningful analytical lens.

Integration Within the MorMag Market Scanner

Within the MorMag Market Scanner, model evaluation extends beyond conventional measures such as return, accuracy, or error metrics. Residual variation is treated as an integral component of the analytical process.

FVU contributes by highlighting the limits of predictive models, signalling when relationships weaken, and providing context for interpreting outputs. It ensures that model signals are not viewed as definitive, but as conditional—dependent on underlying uncertainty and structural dynamics.

This approach reinforces a disciplined framework in which models inform decisions, but do not dictate them.

From Explanation to Awareness

A central insight of FVU is that explanation has limits. In complex systems, not all variation can be captured, and attempts to do so may introduce more harm than clarity.

Recognising this shifts the objective. The goal is no longer to explain everything, but to understand what remains unexplained and to incorporate that uncertainty into decision-making.

This perspective aligns with a broader analytical philosophy: clarity emerges not from eliminating uncertainty, but from acknowledging and navigating it effectively.

Limitations of the Metric

While FVU provides valuable insight, it is not without limitations. Its interpretation depends on model specification, and it does not inherently distinguish between random noise and unmodelled structure. It can also be sensitive to data quality and underlying assumptions.

For these reasons, FVU should be used alongside complementary metrics and interpreted within a broader analytical context. It is a lens, not a definitive measure.

Conclusion

The Fraction of Variance Unexplained offers a powerful perspective on the limits of quantitative models in financial markets. By focusing on what lies outside the model’s explanatory framework, it highlights the presence of uncertainty, noise, and structural complexity.

Within MorMag’s approach, this perspective supports a disciplined methodology in which model outputs are interpreted probabilistically, limitations are explicitly recognised, and uncertainty is integrated into decision-making.

In complex systems, edge does not come from perfect explanation. It arises from understanding the boundary between what can be modelled and what cannot—and operating effectively within that boundary.