This project leverages Google Trends search data as a proxy for consumer behavior indicators to anticipate changes in the Consumer Price Index (CPI).

Data Sources and Preprocessing

Google Trends Search interest data was collected in the United States over a 5-year period for the following keywords:

• gas prices
• rent
• food prices
• used cars
• inflation

Bureau of Labor Statistics (BLS) Monthly CPI data was obtained from the BLS to serve as the ground truth for inflation.

Cleaning & Alignment

• All search trend data was converted to monthly frequency by averaging weekly scores.
• Dates were standardized to “YYYY-MM” format.
• Datasets were merged on the monthly date index.
• Final dataset contained monthly Google Trends scores and corresponding CPI values from May 2020 to April 2025.

Models and Evaluation

The following supervised learning models were applied to predict monthly CPI:

1. Linear Regression $\hat{y} = \beta_0 + \beta_1 x_1 + \cdots + \beta_p x_p$

2. Ridge Regression $\hat{y} = \beta_0 + \sum \beta_j x_j$ with L2 penalty: $\lambda \sum \beta_j^2$

3. Lasso Regression Similar to Ridge but with L1 penalty: $\lambda \sum |\beta_j|$

4. Random Forest Regressor Ensemble model that averages predictions from multiple decision trees.

Train-Test Split: 80/20 Metrics Used:

• Root Mean Squared Error (RMSE)
• Coefficient of Determination (R²)

Model Results & Visualization

1. Best Models: Predicting CPI with Lagged Features

This approach uses one-month-lagged search trend indicators and lagged CPI to predict current CPI levels and mimic real-world forecasting logic.

Performance:

• Linear Regression: RMSE = 0.72, R² = 0.894
• Random Forest: RMSE = 5.74, R² = -5.660
• Ridge: RMSE = 0.73, R² = 0.894
• Lasso: RMSE = 0.71, R² = 0.899

The low RMSE values indicate the model accuracy and the high R² value indicates a strong fit for all the regression-based methods. The random forest model’s performance likely suffered from overfitting.

This approach also accounts for the real-world lag between consumer search behavior and official CPI release and avoids assuming immediate causality.

2. Runner-Up: Predicting CPI % Change with Lagged CPI and Current Trends

This model uses current search trend indicators and lagged CPI to predict current CPI % Change. The rationale was that transforming CPI into a % change makes the target stationary.

Performance

• Linear Regression: RMSE = 0.23, R² = -0.480
• Random Forest: RMSE = 0.32, R² = -1.782
• Ridge: RMSE = 0.23, R² = -0.478
• Lasso: RMSE = 0.21, R² = -0.265

While the low RMSE signals strong accuracy, the negative R² illustrates the poor predictive power of the models. This is likely due to the real-world lag between consumer search behavior and actual immediate economic impact.

Key Insights

• Google Trends data can act as a leading indicator for CPI when preprocessed properly.
• Simpler regression models can provide a strong baseline but fail to adapt to sudden economic shifts.

It is important to recognize that the models don’t account for macroeconomic lags or external shocks. Future work could explore:

• Incorporating additional economic indicators (e.g., unemployment, oil prices)
• Using time series methods like ARIMA or LSTM for sequence modeling
• Feature engineering from text sentiment or regional trends