Simple Feedfoward Neural Networks are Almost All You Need for Time Series Forecasting
This paper shows that simple feedforward neural networks effectively forecast time series, rivaling complex models with improved efficiency and robustness.
What it examines
The paper introduces a simple feedforward neural network (SFNN) for time series forecasting. It aims to show that simple models can match or exceed complex ones like Transformers and GNNs. The study evaluates SFNN designs and critiques current benchmarking practices using various real-world datasets.
What it concludes
The results indicate that SFNNs are effective, robust, and efficient for forecasting. They serve as a strong baseline and can be applied in finance, healthcare, weather, and traffic prediction. The study recommends fairer evaluation protocols and suggests further research into specialized cases where complex models might be needed.
Evidence objects
Research reveals that simple feedforward neural networks match or exceed complex models like Transformers and graph neural networks, showcasing that univariate SFNNs deliver impressive performance with minimal inter-series interaction benefits.
key_findings bullet 1 · key_findings · validation V0
The authors introduce SFNN architecture with a multivariate variant addressing strong interdependencies, execute ablations on components including input mean centering, series-wise mapping, layer normalization, and propose an improved evaluation protocol.
key_findings bullet 2 · key_findings · validation V0
Extensive experiments and cross-validation across multiple datasets reveal SFNNs excel, though performance lags on peculiar datasets such as Traffic, emphasizing simplicity benefits and identifying scenarios where specialized models remain necessary.
key_findings bullet 3 · key_findings · validation V0
Challenging the status quo, this paper reveals that simple feedforward neural networks can rival complex time series forecasting models. It introduces an original perspective on $$\text{model complexity trade-offs}$$, offering a novel, efficient baseline. Its rigorous analysis and critique of benchmarking practices make it highly compelling and impactful in diverse applications.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … Although SFNNs may not be optimal for every situation (hence the “almost” in our title) they serve as a strong baseline that future time series forecasting methods should …
Source row: 1781 · abstract type: snippet