Proposal Description

This session brings together four papers showcasing how contemporary forecasting methods are being applied across economic and financial domains. The contributions highlight how traditional statistical tools, adaptive mathematical frameworks, and modern machine learning techniques can be combined to improve predictive accuracy and support policy analysis.

The first paper demonstrates how web-scraped supermarket prices can be used to nowcast food inflation, where the food component accounts for a large percentage of the Harmonised Index of Consumer Prices (HICP). Using a newly assembled dataset of various products and daily prices from supermarkets and corner shops; classified through string-matching techniques and large language models, the study evaluates three forecasting strategies: a naïve average across products, a minimum-distance approach aligned with official statistics, and a machine learning framework based on mixed-frequency regressions. Out-of-sample comparisons with the Narrow Inflation Projection Exercise (NIPE) indicate that online prices can meaningfully enhance short-term prediction, though performance varies across food categories. The results remain preliminary but underscore the potential of high-frequency digital data for real-time inflation monitoring.

The second paper addresses medium- and long-term demographic forecasting, focusing on ageing population and labour-force dependency ratio through 2050. By comparing ARIMA models with the Prophet algorithm, the study examines how different techniques capture demographic trends and structural shifts. Given the implications for pension sustainability, healthcare planning, and labour-market policy—challenges shared by many European countries—the analysis highlights the importance of selecting forecasting methods capable of handling gradual but consequential demographic changes.

The third paper introduces an information-theoretic extension to spline smoothing designed for financial time series. Building on reproducing kernel Hilbert space theory, the proposed framework explicitly incorporates how information arrives and propagates through markets while preserving temporal causality and accommodating time-varying volatility. This approach provides new avenues for modelling volatility surfaces and supports applications in portfolio optimisation and risk management that better reflect the informational structure of financial markets.

The final paper examines the use of deep learning in the production of economic statistics, with a focus on recurrent neural networks applied to monthly external trade figures. Beyond evaluating predictive performance, the study considers operational questions relevant to central banks and statistical institutions, including integration into existing workflows, computational requirements, and the broader implications of adopting artificial intelligence in official statistics. The findings contribute to ongoing discussions about how advanced modelling techniques can complement established methodologies in economic data production.

Collectively, these papers illustrate how effective forecasting increasingly relies on the blending of classical statistical methods, adaptive mathematical frameworks, and artificial intelligence. Each approach offers unique strengths for specific forecasting challenges, and their thoughtful integration supports more accurate projections and better-informed decision-making across domains ranging from inflation and demography to financial markets and economic statistics.