『Executive Summary』

     Two main sources of data are typically considered to forecast inflation. One source, namely the market-based data, is from real activity measures based on the Phillips model, using information embedded in the interest rates and asset prices. Another source comes from surveys, where questions related to future inflation are asked to the professionals or consumers. The limitation of the former is usually associated with relatively unstable and limited forecasting accuracy, while the latter is pointed out for requiring heavy costs to conduct, bias, and limited sample ranges.

     Under such context, this project (1) attempts to improve the forecasting power of the survey-based forecast via implementing bias adjustments, and (2) introduces the third alternative method of inflation forecasting as a complementary measure, which exhibits plausible potentials to be the next-generation forecasting method with less cost and supplementary accuracy: the text sentiment analysis via a designed lexical corpus, utilizing big data from the media sources and bias-adjusted survey forecasts.

     Accordingly, this project examines the predictability of alternative forecast methods derived from inflation surveys and news sentiment for three inflation measures: CPI all-items, core CPI, and all-items-less-shelter CPI. The first part of this paper constructs a series of sentiment indices from over 14,000 news articles about the U.S. inflation in New York Times between 1978 to 2014. The news sentiment is measured by lexical methodology with a professional dictionary in the economic domain published by Loughran and McDonald. In terms of survey-based forecasts, this project adjusts bias for SPF and MICH. The empirical findings from the generated model suggests that the bias adjustment of the survey data depicts greatly improved accuracy at forecasting CPI inflation. The individual forecast constructed by news sentiment shows contemporaneous explanatory power regarding inflation, but relatively limited predictive content. Nonetheless the designed combination model over the full sample displays an outstanding accuracy, with news sentiment functioning as an effective complementary source to surveys particularly at forecasting recent periods.

     These results thus imply three potential attributes regarding bias-adjustments in surveys and media-based text sentiment measures using lexical methodology. First, econometric technology for bias adjustment is a must for survey-based methods, as it generates significantly improved forecasting power. Second, the constructed sentiment indices can be used as an effective complementary measure for accurate information forecasting, with relatively less cost compared to the market-based or survey-based measures. Third, the ongoing development of machine learning technology implies further space for media-based text sentiment measure to be the next-generation source of inflation forecasting, with potentials to replace the conventional approaches under the superior resource-saving efficiency and broader sample ranges.

Table of Contents                                                          

Executive Summary................................................................................................1

1. Introduction.........................................................................................................2

2. Research Proceedings on Inflation Forecasts......................................4

3. Data Configuration..........................................................................................7

4. Methodology and Forecasting Model Formulations.....................15

5. Results and Interpretations.......................................................................25

6. Conclusion.......................................................................................................35

References............................................................................................................37