Abstract
This study investigates the association between monthly sunshine, consumer confidence and retail volume in Germany, Spain, Türkiye and the United Kingdom. We use the official OECD consumer-confidence series and evaluate fixed models using six expanding calendar-year windows with a one-month embargo. All imputation, scaling and PCA are fitted within training folds. Country-level sunshine regressions use seasonal and trend controls, HAC(12) covariance, geography/difference sensitivity and a 12-test false-discovery adjustment. No specification passes the 5% adjusted threshold. Persistence is a strong CCI benchmark (RMSE 0.492); a combined weather–macro forest achieves retail R² 0.420, with little difference from its macro-only counterpart. The results do not establish a causal weather effect or a deployable real-time forecasting advantage.
1. Research question
My interest in psychology and decision economics led me to ask whether environmental conditions are associated with economic sentiment and spending. The study examines monthly sunshine and consumer confidence, compares weather and macroeconomic predictors, and evaluates their predictive performance against simple reference models. Consumer confidence is measured using the official OECD CCICP series.
2. Data and design
The balanced frame contains 480 country-months. CCI covers all observations; retail is observed in Germany, Spain and Türkiye only (360 rows). The main weather measure uses a five-city population-weighted weather snapshot, with the capital-coordinate snapshot used for sensitivity. Missing macro controls are preserved until training-fold imputation. No target is filled. The frozen inputs, exact OECD selection and SHA-256 fingerprints are public.
3. Methods
For each test year from 2019 through 2024, training stops before the preceding December. Each preprocessing transformer is fitted on that training subset. Models include country-mean and one-month persistence baselines, weather/macro/combined forests, a lag-augmented forest and PCA-Ridge. The direction task uses only well-defined next-month labels and purges training labels at the test boundary. The statistical specifications control for calendar month and trend, estimate HAC(12) intervals, and adjust the 12 specified tests using Benjamini–Hochberg.
4. Findings
Every adjusted q-value exceeds 0.05. Effects vary by geography and level/difference specification, so the study provides no robust evidence for a common sunshine effect on sentiment. This is uncertainty, not proof of absence. Persistence outperforms the non-autoregressive CCI models. For retail, macro-only and combined forests perform similarly; a strong combined score cannot be attributed to weather alone. Full annual and country metrics remain visible, including negative R² values.
5. Limits
City aggregation can mismeasure household exposure. Fixed city weights lack a recorded population reference year. The macro/retail source response vintage was not retained. Economic releases are revised; same-month predictors may be unavailable at a practical forecast origin. Annual fold errors are dependent, and no “significant model improvement” claim is made. HAC and multiple-testing corrections do not address every omitted variable, structural break or nonlinear relationship. Only four countries and a decade are studied.
6. Conclusion
The study combines country-level weather and economic data with out-of-time evaluation, explicit data coverage and interpretable reference models. The evidence supports caution about broad weather–sentiment claims and makes the project reproducible enough to extend with higher-resolution exposure and real-time data vintages.