Relationships Among Economic Variables using Machine Learning Techniques
Geoff Lee| 11/4/2023
Australia’s National Science Agency
Background • Strategic decisions (investment, policy, etc) that will have future impacts need a forecast of that future to determine appropriate action • Quantitative models provide a consistent and rigorous way to generate correlated forecasts – uncertainty, sensitivity • Useful for cost-benefit analysis • Understanding relationships among variables is crucial when establishing a multi-variate forecast model • Typically, theoretical models are calibrated to historical data using regressions, then used to forecast economic conditions into the future Background • As decision time horizon increases, models become less useful: • Models’ high uncertainty dominates • Relationships between economic variables may change over the longer horizon • Theoretical models and assumptions may not hold in practice • Decision-makers become less interested in absolute value, more interested in relative levels – rely more on qualitative future scenarios (e.g. we think the electricity price will double in 10 years) and somewhat independently risk factor to risk factor • Machine learning techniques offer a complementary approach – uncover relationships that may not be apparent Methodology • Clustering techniques commonly used to analyse financial data – mostly in stocks/economic performance • Use shape-based approach to cluster time series to explore relationships between variables more broadly • Complexity-Invariant Distance measure for clustering metric – corrects a standard distance measure such as Euclidean distance – better results than DTW Methodology
• Global kernel k-means algorithm – the input space mapped to
kernel function – not dependent on initial condition and avoids local minima in finding clusters • Kernel functions can include polynomial, Gaussian radial basis functions and sigmoid • The elbow method Case Study Data • Choose three data sets to explore across different domains • Agriculture: Beef price • Energy: Natural gas price • Economic: Australian Consumer Price Index (CPI) rate • Data sources: • Monthly beef price from Bloomberg (Jan 1993 – Dec 2022) • Daily trading spot price from the Henry Hub Natural Gas price (Jan 1993 to Sept 2021) • Quarterly CPI data from the Australian Bureau of Statistics (1 st qrtr 1993 – 1st qrtr 2023) • Take quarterly data from Jan 1993 to Sept 2021. Data Procedure • Choice of lag: Procedure Results Results Results Conclusions
• Novel clustering methodology for analysing time series data and
identifying relationships between economic variables • Methodology demonstrated through case study with three variables – gas price, beef price and CPI • Web visualisation tool for decision makers to interact
• Future work – causality, identifying links between variables