A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition (1108.3259v1)

Abstract: Multi-step ahead forecasting is still an open challenge in time series forecasting. Several approaches that deal with this complex problem have been proposed in the literature but an extensive comparison on a large number of tasks is still missing. This paper aims to fill this gap by reviewing existing strategies for multi-step ahead forecasting and comparing them in theoretical and practical terms. To attain such an objective, we performed a large scale comparison of these different strategies using a large experimental benchmark (namely the 111 series from the NN5 forecasting competition). In addition, we considered the effects of deseasonalization, input variable selection, and forecast combination on these strategies and on multi-step ahead forecasting at large. The following three findings appear to be consistently supported by the experimental results: Multiple-Output strategies are the best performing approaches, deseasonalization leads to uniformly improved forecast accuracy, and input selection is more effective when performed in conjunction with deseasonalization.

• The paper systematically compares five forecasting strategies, revealing that Multiple-Output methods best capture dependencies between future data points.
• The paper demonstrates that deseasonalization consistently enhances forecasting accuracy, especially when combined with effective input selection.
• The empirical analysis employs rigorous statistical tests to validate findings, providing actionable insights for advancing multi-step time series forecasting.

Analysis of Multi-Step Ahead Time Series Forecasting Strategies

The paper "A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition" presents an in-depth analysis of various methodologies used to tackle the complex problem of forecasting time series multiple steps into the future. This research leverages the challenging and realistic dataset from the NN5 forecasting competition, which involves diverse series fraught with issues such as missing data points and prominent seasonal patterns.

Core Contributions

The paper identifies and systematically compares five fundamental strategies for multi-step ahead forecasting:

1. Recursive Strategy: Utilizes a single model iteratively for forecasting each step ahead. Despite its simplicity, it is prone to error accumulation.
2. Direct Strategy: Involves multiple models, each designed to predict a specific future point. This approach avoids error propagation but assumes independence between forecasts, which can be limiting.
3. DirRec Strategy: Combines the strengths of the Recursive and Direct methods, using different models for each step while feeding back forecasts as inputs.
4. MIMO Strategy: A Multi-Input Multi-Output strategy that predicts all future steps in one go, preserving inter-dependencies but at the cost of flexibility in model structure.
5. DIRMO Strategy: Offers a balance between dependency preservation and modeling flexibility by splitting forecasts into intermediate blocks.

Experimental Evaluation

The authors perform extensive empirical tests using the NN5 dataset, consisting of 111 time series, each representing ATM withdrawal patterns in the UK. The experimental setup is rigorous, with considerations for deseasonalization, input selection, and multiple model combinations. These are tested under various configurations to validate the performance of each strategy.

Three main findings are highlighted:

• Multiple-Output approaches (MIMO and DIRMO) consistently outperform Single-Output approaches by capturing dependencies between future data points.
• Deseasonalization uniformly enhances forecasting accuracy, reinforcing its importance in handling time series with seasonal variations.
• Input selection is notably more effective when coupled with deseasonalization, suggesting a synergetic effect between the two preprocessing steps.

Methodological Rigor

The results are validated using the Friedman and Iman-Davenport statistical tests to ensure significance across multiple datasets. Post-hoc tests further dissect pairwise differences between strategies to offer robust insights.

Implications and Future Research

The findings underscore the efficacy of Multiple-Output strategies and the critical role of deseasonalization. While Recursive and Direct methods have their own merits, they often fall short in complex scenarios involving significant temporal dependencies. The results advocate for future research to refine Multiple-Output strategies further, perhaps by incorporating more sophisticated deseasonalization techniques tailored to these methods.

This paper provides valuable guidance for both theoretical exploration and practical application in time series forecasting, presenting quantitative benchmarks and methodological insights that can influence future advancements in AI and machine learning-based forecasting.