In this post, I use machine learning and real macroeconomic indicators to analyze the drop in consumer confidence during Q1 2025. The goal: determine if the decline was justified by fundamentals — or driven by public sentiment and perception.
Figure: Updated model predictions vs actual Consumer Confidence Index showing a clear divergence in Q1 2025
📉 Investigating the Consumer Confidence Drop in Q1 2025
🎯 Objective
In early 2025, consumer confidence dropped sharply across the United States. I wanted to know: Was this decline justified by the underlying economy? Or did public sentiment fall further than the data would predict?
🔍 Approach
- ✅ Built a machine learning model trained on seven years of U.S. economic indicators
- ✅ Predicted what consumer confidence should have been based on those fundamentals
- ✅ Compared the model’s predictions to actual confidence levels over two periods:
- Q4 2024 (Oct–Dec): A period of relative stability
- Q1 2025 (Jan–Mar): A surprising decline in sentiment
📊 Results
✅ Model Performance
| Period | MAE | RMSE | MAPE |
|---|---|---|---|
| Q4 2024 | 0.01 | 0.00 | 0.01% ✅ |
| Q1 2025 | 0.74 | 0.92 | 1.14% ❌ |
The model predicted Q4 nearly perfectly, but showed a noticeable decline in accuracy during Q1 — suggesting sentiment dropped below expectations.
🧪 Statistical Testing
To determine whether this decline was statistically meaningful, I compared model residuals for Q4 vs Q1 using three standard tests:
| Test | T / U Statistic | p-value | Interpretation |
|---|---|---|---|
| Welch’s t-test | 10.67 | < 0.0001 | Mean errors are significantly different |
| Mann-Whitney U | 5639.00 | < 0.0001 | Distributions of residuals are not the same |
| Levene’s Test (variance) | 49.78 | < 0.0001 | Variance in Q1 errors is much higher |
✅ All results are statistically significant, confirming that the model failed to account for sentiment shifts in Q1.
🔍 Visual Evidence: Actual vs Predicted
Below is a comparison of actual vs predicted consumer confidence values in each quarter. The model performed almost perfectly in Q4 — closely tracking actual values — but began to overestimate confidence in Q1 as sentiment dropped below expected levels.
Figure 1: The model tracks consumer confidence closely in Q4 2024, but begins to overestimate it in Q1 2025 as sentiment drops below expected levels.
📉 Visual Evidence: Residual Plots
The plots below show the residuals (prediction errors) over time. In Q4 2024, residuals are extremely small, hovering close to zero, with y-axis ticks as tight as ±0.02. This reflects the model’s near-perfect accuracy in that period.
In contrast, Q1 2025 shows a much wider residual range, with consistent negative errors — meaning the model overpredicted confidence. This aligns with our hypothesis that sentiment fell unexpectedly in early 2025.
Figure 2: Residual error plots reveal the full story. Q4 residuals are tiny (note the y-axis scale), indicating strong alignment between model and reality. In Q1, errors are consistently negative — showing that the model consistently overestimated confidence.