Kyle Polich avatar

Dexa/Kyle Polich

Learn more
Published Oct 17, 2021

Fast and Frugal Time Series Forecasting

Explore the future of time series forecasting with Fotios Petropoulos, as he delves into collaborative research, innovative methodologies, and the balance between statistical and managerial insights. Discover strategies to optimize computational costs and weigh simplicity against complexity in model selection for accurate and efficient predictions, particularly in retail contexts.
Episode Highlights
Data Skeptic logo

Topics covered

Popular Clips

Episode Highlights

  • Model Selection

    Selecting the right forecasting model involves balancing statistical measures and practical considerations. emphasizes using corrected AIC for small sample sizes and introduces a novel approach called forecasting by representativeness. This method compares out-of-sample forecasts with in-sample data to avoid explosive forecasts from models like multiplicative trend models 1. He also highlights the importance of simplicity, suggesting a reduced set of eight exponential smoothing models to cover various forecasting profiles 2.

    We try to reduce the number of models through which our algorithm goes, instead of going through all 30 models of exponential smoothing.

    ---

    This approach aims to streamline model selection while maintaining accuracy.

       

    Statistical vs ML

    The debate between statistical and machine learning methods in forecasting is ongoing. notes that while machine learning methods like LightGBM have shown promise, especially in higher aggregation levels, their advantage diminishes at more granular levels 3. He stresses the importance of considering the cost of forecast errors and suggests measuring the utility of forecasts in practical applications, such as inventory management 4.

    The outperformance of machine learning over statistical methods is not consistent across different hierarchical levels.

    ---

    This highlights the need for a balanced approach in model selection.

       

    Model Simplicity

    Simplicity in forecasting models can often lead to better results. argues that a smaller set of models can effectively capture essential time series patterns like trend and seasonality 5. By reducing the number of models, the search for optimality decreases, allowing for good enough forecasts without the computational burden of complex models 2.

    Simple models could produce good results, even better results from more complicated models.

    ---

    This approach aligns with the principle of Occam's razor, favoring simpler solutions.

Related Episodes