Navigating Election Forecasting: Why Uncertainty Often Outweighs the Shock

Introduction: The Challenge of Predicting Local Elections

Forecasting English local elections is notoriously difficult. Unlike national contests, local polls involve thousands of wards, shifting boundaries, and low turnout patterns that amplify randomness. Traditional forecasting models often fail because they try to predict a single outcome when the real world is defined by uncertainty. This article explores how scenario modelling—a technique that embraces ambiguity rather than fighting it—can provide more robust insights, especially when historical errors and calibrated uncertainty reveal that some models are most valuable precisely when they refuse to produce a single forecast.

Understanding Calibrated Uncertainty

What Is Calibrated Uncertainty?

Calibrated uncertainty refers to a model's ability to accurately communicate the range of possible outcomes, not just a point estimate. In election forecasting, this means providing probability distributions rather than a single predicted vote share. For English local elections, where data granularity is low and idiosyncratic factors (like local issues or candidate quality) dominate, calibrated uncertainty helps analysts avoid overconfidence. For example, a well-calibrated model might say a party has a 40–60% chance of winning a council, with a wide credible interval, instead of stating exactly 52%.

Why Calibration Matters More Than Accuracy

A model that is perfectly calibrated—meaning its predicted probabilities match observed frequencies over many events—is more trustworthy than one that occasionally lands on the exact number but is often wrong about its own reliability. In the context of local elections, a model that refuses to forecast a winner because uncertainty is too large is often more honest and useful than one that produces a confident but misleading prediction. This is especially true when the uncertainty is bigger than the shock—that is, when the inherent variance in outcomes dwarfs any single event or trend.

The Role of Historical Error

Learning from Past Mistakes

Historical error analysis is central to building good scenario models. By examining how previous forecasts deviated from actual results, modellers can quantify systematic biases and random noise. For English local elections, historical errors often stem from poor turnout models, incorrect boundary changes, or unanticipated national swings. Scenario modelling uses these error distributions to generate thousands of plausible futures, each weighted by how likely it is based on past performance. This approach turns hindsight into a tool for foresight.

Error Asymmetry and Its Implications

Not all errors are equal. A model that consistently overestimates Labour support in suburban wards has a different error profile than one that underestimates Conservative strength in rural areas. Scenario analysis can highlight these asymmetries, allowing campaigners to adjust strategies. For instance, if historical error shows that a model's predictions for swing seats are twice as variable as for safe seats, then the uncertainty around those races should be reflected in the scenarios. This nuanced view prevents the false precision that plagues many election forecasts.

When Models Refuse to Forecast

The Paradox of Predictive Refusal

Some of the most useful models are those that explicitly declare they cannot make a meaningful prediction. This might happen when input data is too sparse, the situation is too volatile, or the uncertainty intervals overlap so broadly that no single outcome stands out. In such cases, a model that refuses to forecast is providing critical information: that the system is fundamentally unpredictable under current conditions. For decision-makers, this is far better than a false sense of certainty. It forces them to prepare for a wide range of possibilities rather than betting on a single scenario.

Practical Examples in English Local Elections

Consider a ward where an independent candidate with no past electoral history enters the race. Historical polling data may be entirely insufficient. A responsible scenario model would generate a wide confidence interval, perhaps ranging from 5% to 55% vote share. Rather than forcing a central estimate, the model can output probability distributions and note that any single forecast would be misleading. This refusal is a strength: it tells the candidate and party staff that they need to gather more data or accept that the outcome is highly contingent on local factors.

Building a Scenario Model for Local Elections

Step 1: Gather and Calibrate Inputs

Start with historical vote shares, turnout data, and demographic trends. Use Bayesian methods to combine prior beliefs with observed data, producing posterior distributions that automatically account for uncertainty. Calibrate the model by comparing past predictions with actual results, adjusting for systematic biases.

Step 2: Generate Multiple Scenarios

Create thousands of simulation runs, each drawing from the error distributions and input uncertainties. For English local elections, this might include variations in national swing, local campaign effectiveness, and weather on election day (which affects turnout). Each scenario represents a plausible reality.

Step 3: Communicate Uncertainty

Present results as ranges or fan charts, not single numbers. Highlight where the model's confidence is high (e.g., safe seats) and where it is low (e.g., marginal contests with high historical error). For scenarios where uncertainty dominates, label them clearly and explain why a forecast is not possible.

Conclusion: Embracing Uncertainty as a Feature

English local elections are inherently uncertain, and pretending otherwise does a disservice to voters, candidates, and analysts. Scenario modelling that incorporates calibrated uncertainty and historical error offers a path forward, turning the model's refusal to deliver a single forecast into a valuable insight. When the uncertainty is bigger than the shock, the most honest model is often the one that says, “I don’t know—and here is exactly why.” By focusing on ranges, probabilities, and the limits of prediction, decision-makers can navigate the fog of election night with eyes wide open.

Learn more about calibrated uncertainty and how it applies to other forecasting domains.