Impact of Internet Search Trends on Predicting Localized Economic Indicators Using Time Series and Bayesian Structural Time Series Models

Project Abstract

This study investigates the predictive power of internet search trends for localized economic indicators using time series analysis and Bayesian Structural Time Series (BSTS) models. By leveraging high-frequency search data from major search engines alongside traditional macroeconomic series, we examine whether search activity can improve nowcasting and short-horizon forecasting across diverse urban and regional contexts. The research employs a multi-stage methodology (i) data collection and preprocessing, including alignment of daily search volumes with monthly or quarterly economic indicators such as unemployment claims, consumer confidence, retail sales, and small business sentiment at fine geographic granularity; (ii) feature engineering to capture trend, seasonality, and exogenous shocks within search data, incorporating lag structures and Google Trends-derived proxies for interest in local economic activities; (iii) specification of BSTS models that decompose time series into trend, seasonal, regressors, and error components, with informative priors to handle sparse regional data and to quantify structural breaks; (iv) model comparison against conventional benchmarks—ARIMA, VAR, and machine learning approaches—using out-of-sample predictive accuracy, nowcasting timeliness, and probabilistic forecast intervals; and (v) robustness checks for data quality issues, endogeneity, and non-stationarity. We further extend BSTS with hierarchical and regional pooling to capture cross-area dependencies while preserving local specificity. The study addresses key research questions the extent to which search trends add predictive value above and beyond traditional indicators, the optimal lags and combinations of search-derived signals for different economic domains, and the conditions under which BSTS yields superior stochastic forecasts with credible intervals in granular markets. Empirical results demonstrate that incorporating search trends improves short-term forecast accuracy for several localized indicators, particularly in periods of economic inflection or abrupt policy changes. The hierarchical BSTS framework reveals heterogeneity in signal strength across regions, with urban areas exhibiting stronger alignment between search activity and employment-related outcomes, while rural regions benefit more from interaction terms and exogenous covariates. The probabilistic nature of BSTS enables improved uncertainty quantification, producing calibrated prediction intervals that adapt to regime shifts and data revisions. Sensitivity analyses indicate that the benefits are contingent on data quality, filtering of noise in search signals, and careful handling of multiplicity in regional models. The findings have practical implications for policymakers, regional planners, and financial analysts by providing a timely, data-driven approach to monitor localized economic conditions, allocate resources efficiently, and anticipate turning points. The study also discusses methodological limitations, including potential biases in search behavior, data privacy considerations, and the challenge of real-time data availability, offering avenues for future work such as integrating alternative digital traces, exploring non-linear BSTS extensions, and expanding coverage to a broader set of regional economies.

Project Overview

What This Project Is About

---

A plain-language overview of how people’s online search behavior can help predict local economic conditions. The project looks at whether patterns in search queries relate to indicators like unemployment, consumer spending, and business activity in a specific area, using simple time-based analyses and more advanced statistical models.

---

---

The Problem It Addresses

---

Forecasting local economic trends can be noisy or lagging when using traditional data alone. This project investigates whether real-time search data can provide timely signals to improve predictions, addressing gaps in early warning and resource planning for communities and policymakers.

---

---

Objectives of the Project

---

1. Understand what data types are relevant (search trends and economic indicators).
2. Explore simple ways to relate search activity to local indicators.
3. Assess whether search data adds value beyond traditional methods.
4. Learn basic time-series concepts and a beginner-friendly Bayesian approach.

---

---

What You Will Do Step by Step

---

1) Gather publicly available search trend data for a chosen locality and corresponding local economic indicators. 2) Clean and organize data into a common timeline. 3) Perform descriptive analyses to spot apparent links. 4) Apply straightforward time-series methods to test predictive power. 5) Introduce a simple Bayesian form of the model to handle uncertainty. 6) Compare predictions with and without search data. 7) Interpret results in clear, non-technical terms. 8) Reflect on limitations and practical implications.

---

---

---

---

Expected Outcome

Clear findings on whether online search trends can improve short-term predictions of local economic indicators, with an approachable explanation of how much value the data adds and where it may fall short. The project should yield practical guidance for researchers and policymakers on using accessible data sources for local forecasting.