Investor attention and consumer price index inflation rate: Evidence from the United States

The paper investigates whether investor attention, an emerging concept in behavioral finance, helps explain and forecast inflation in the United States. Inflation remains a central macroeconomic concern, especially amid recent shocks such as COVID-19, monetary expansions, carbon neutrality expectations, and oil price volatility. Existing research emphasizes traditional determinants (interest rates, oil prices) and models (e.g., NKPC), but forecast accuracy has plateaued and novel drivers are underexplored. The authors argue that investor attention is theoretically linked to inflation because attention affects asset pricing and information acquisition, which shapes inflation expectations—an NKPC determinant of actual inflation. The study formulates three hypotheses: H1: Investor attention can empirically explain inflation; H2: Investor attention can empirically forecast inflation; H3: Investor attention affects inflation through its influence on inflation expectations. The paper connects micro-level investor behavior to macro-level inflation dynamics to fill this research gap.

Prior work explains inflation via monetary policy and Phillips curve frameworks, and attributes inflation dynamics to factors like industrial production, unemployment, wages, aggregation mechanisms, exchange rates, firm costs, and supply chain pressures. Forecasting methods include VARs, dynamic factor models, neural networks, Bayesian methods, and survey-based forecasts; recent advances leverage machine learning across various countries. Novel predictors such as commodity price aggregates, metal and oil price information, online price indices, behavioral measures (investor sentiment), agricultural prices, climate variables, and carbon market returns have improved inflation forecasts to varying degrees. Within behavioral finance, investor sentiment and attention have been linked to asset pricing, FX, commodities, carbon, and cryptocurrency markets, and to ESG performance and volatility spillovers. However, despite suggestive connections, no study has directly examined investor attention’s role in inflation, and sentiment-based findings for inflation are mixed. This gap motivates examining investor attention as a determinant of inflation and as a predictor for inflation forecasts.

Data: Investor attention is proxied by the Google Search Volume Index (GSVI) from Google Trends, using the keyword “inflation,” United States region, monthly frequency from January 2004 to July 2020. Inflation is measured by seasonally adjusted CPI for all urban consumers (CPIAUCSL), transformed to monthly log differences. Both series are differenced in logs and denoted Att and Inf. Stationarity is verified using ADF, KPSS, and PP tests, confirming suitability for VAR modeling. The sample is split into in-sample (Jan 2004–Aug 2016) for model estimation/explanation and out-of-sample (Sep 2016–Jul 2020) for forecasting. Models: - VAR and Granger causality: A bivariate VAR with lag length selected as 1 captures dynamics between Inf and Att. Granger causality is tested via joint significance of lagged terms using chi-square statistics. Stability is validated via AR root tests and Impulse Response Functions (IRFs) evaluate the response of inflation to attention shocks. - Linear regression with controls: To account for key external influences, Brent oil futures returns (oil) are included, and interactions between attention and oil are modeled: Inf_t = α0 + Σ α_i Inf_{t−i} + Σ β_i Att_{t−i} + Σ γ_i oil_{t−i} + Σ ρ_i (Att_{t−i} × oil_{t−i}) + ε_t. Forecasting setup: Out-of-sample forecasting models extend the explanatory frameworks with lead-lag structure to predict Inf_t one-, two-, and three-months ahead using rolling windows and fixed estimation window sizes. Two specifications are used: (i) Inf_t regressed on lags of Inf and Att; (ii) the same plus oil lags and interaction terms. Benchmarks are AR(1) and Random Walk (RW). Forecast accuracy is evaluated via out-of-sample R^2, MSFE, and MSFE-adjusted statistics (Clark-McCracken) to test significance of forecast improvements. Robustness checks: 1) Alternative attention proxies via Google Trends: “monetary policy” alone and a combined measure summing searches for “inflation” and “monetary policy.” The full empirical pipeline (explanation, forecasting) is repeated. 2) Alternative model specifications and variables: Add real interest rate (FRED) as an additional control; include higher moments (squared terms) for Inf and Att; and consider interaction-based and regime-dependent models following prior literature. Specifications include: - Inf_t = α0 + α1 Inf_{t−1} + β1 Att_{t−1} + γ0 oil_{t−1} + δ1 Rate_{t−1} + ε_t; - with interactions Att×oil and Att×Rate; - higher-moment model including Inf_{t−1}^2 and Att_{t−1}^2; - interaction of Inf_{t−1}×Att_{t−1}; - and interaction with a regime indicator D(Inf_{t−1} ≤ 0). Mechanism test (expectations channel): To test H3, the paper regresses University of Michigan survey-based inflation expectations (Exp_t) on 12 monthly lags of Att: Exp_t = c + Σ_{i=1}^{12} β_i Att_{t−i} + ε_t, motivated by information rigidities and delayed processing. Variance Inflation Factor (VIF) tests assess multicollinearity across lagged Att regressors.

- Stationarity and dynamics: Both Inf and Att are stationary in monthly log-differences. VAR(1) stability is confirmed (roots inside unit circle). IRF shows a one-unit shock to investor attention affects inflation for roughly seven months. - Explanatory power and causality: In VAR(1), Att_{t−1} significantly and negatively predicts Inf_t (coefficient ≈ −0.0038, p<0.05), while Inf does not Granger-cause Att. Granger test shows investor attention Granger-causes inflation (χ² ≈ 3.8877, p<0.05). With oil controls and interactions, attention remains significantly negative (Att_{t−1} ≈ −0.0033, p<0.05), oil_{t−1} positive (≈ 0.0210, p<0.01), interaction positive (≈ 0.0362, p<0.05), R² ≈ 0.525. - Short-horizon forecasting (1-month ahead): Models including Att outperform AR(1) and RW benchmarks. For example, relative to AR(1), out-of-sample R² values are positive (e.g., ≈ 0.059–0.183) with significant MSFE-adjusted statistics; relative to RW, R² values are large (e.g., ≈ 0.737–0.799) with significant MSFE-adjusted statistics, indicating notable gains from including attention and oil/interaction terms. - Long-horizon forecasting (2–3 months ahead): Predictive models with Att continue to beat both benchmarks. For horizon=2 vs AR(1): R² ≈ 0.0620 and 0.2306; MSFE-adjusted ≈ 1.98**. Versus RW: R² ≈ 0.9158 and 0.9670; MSFE-adjusted ≈ 4.48*** and 6.20***. For horizon=3 vs AR(1): R² ≈ 0.0705 and 0.1692; MSFE-adjusted ≈ 1.96–2.03**. Versus RW: R² ≈ 0.9525 and 0.9666; MSFE-adjusted ≈ 5.19–5.68***. - Robustness: Using alternative keywords, investor attention remains a significant negative predictor and Granger cause of inflation. Forecast gains persist when the attention proxy aggregates “inflation” and “monetary policy” (notably strong against RW; improvements vary vs AR(1)). Alternative specifications—adding real interest rate, higher moments, and interaction/regime models—consistently show a significantly negative Att coefficient. - Mechanism via inflation expectations: Distributed-lag regressions of Michigan survey expectations on Att_{t−i} (i=1,…,12) show all lag coefficients positive and significant at 1%. VIF values around ~2.1–2.3 indicate no serious multicollinearity concerns. This supports H3: investor attention influences inflation through its effect on inflation expectations. Overall: Investor attention negatively affects current inflation, Granger-causes inflation, and materially improves out-of-sample inflation forecasts at both short and long horizons. Results are robust to alternative attention measures and model specifications, and the expectations channel is empirically supported.

The findings directly address the research questions and hypotheses. Investor attention, proxied by Google search intensity, contains leading information about inflation dynamics: it Granger-causes and negatively predicts CPI inflation. The negative relationship is consistent with mechanisms where heightened attention to inflation prompts intertemporal shifts in consumption and investment behavior and aligns with limited-attention/recognition hypotheses that associate increased attention with lower subsequent returns and prices. Forecasting results demonstrate that integrating attention meaningfully improves predictive performance relative to simple benchmarks across multiple horizons, highlighting the practical value of behavioral indicators in macroeconomic forecasting. The robustness analyses underscore that the relationship is not sensitive to reasonable variations in attention measurement or model specification. Finally, the documented linkage from attention to survey-based inflation expectations provides a plausible transmission channel consistent with NKPC frameworks: attention facilitates information acquisition, shaping expectations that, in turn, influence realized inflation. These insights suggest that investor/public attention is a relevant state variable for macroeconomic monitoring and policy communication strategies.

This paper establishes investor attention as a significant determinant and predictor of U.S. CPI inflation. Empirically, attention negatively impacts inflation and Granger-causes it; incorporating attention into forecast models yields substantial out-of-sample gains at short and long horizons versus AR(1) and RW benchmarks. The results are robust to alternative attention proxies, additional controls (oil, real rates), higher-moment and interaction specifications. Moreover, attention significantly influences inflation expectations, supporting an expectations-based transmission mechanism. Contributions include introducing investor attention into inflation analysis, demonstrating its explanatory and predictive value, and linking attention to expectations in a macro context. Future research should: (i) explore more sophisticated, potentially nonlinear and regime-switching models to capture complex dynamics; (ii) examine alternative, possibly cumulative measures of attention and richer information structures for expectations formation; and (iii) evaluate external validity across countries, inflation regimes, and higher-frequency settings.

The study acknowledges two primary limitations: (1) reliance on linear specifications may not fully capture potentially nonlinear or regime-dependent dynamics between attention and inflation, suggesting the need for more sophisticated models; (2) the expectations–attention linkage is modeled linearly, whereas information processing may be more complex; exploring alternative models and cumulative attention measures is warranted. Additionally, while robustness checks cover alternative keywords and model specifications, the analysis focuses on U.S. monthly data from 2004–2020 and specific benchmarks (AR and RW), which may limit generalizability to other periods, countries, or benchmark frameworks.