Research map / Time series / Long-term forecasting
FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FiLM's Legendre projection and Fourier-enhanced architecture instead of FEDformer | Long-term forecasting · Six multivariate benchmarks, all reported horizons | MSE: 20.3% | Helped |
| Use FiLM instead of FEDformer | Long-term forecasting · Six univariate benchmarks | MSE: 22.6% | Helped |
| Add Legendre projection in place of a comparable-sized linear layer | MLP, LSTM, CNN, and Attention forecasting backbones · ETTm1 and Electricity | MSE: MLP: 119.4%; LSTM: 97.0%; CNN: 13.8%; Attention: 8.2% | Helped |
| Replace LPU with comparable-sized linear layers (ablation) | FiLM · ETTm1 and Electricity | MSE: 32.5% degradation on average | Hurt |
| Replace FEL with MLP (ablation) | LPU-based forecasting model · ETTm1 and Electricity | MSE: +0.0%; -1.5%; -1.7%; -0.9%; +155%; +59%; +55%; +33% | Mixed |
| Replace FEL with CNN (ablation) | LPU-based forecasting model · ETTm1 and Electricity | MSE: +13.5%; +10.8%; +10.8%; +8.6%; +330%; +224%; +128%; +192% | Hurt |
| Apply low-rank approximation to FEL weights (ablation) | FiLM · ETTh1 and Weather | MSE: Parameter size: 100% to 0.4% | Mixed |
| Replace lowest-frequency mode selection with random frequency selection (ablation) | FEL · Exchange and Weather | MSE: Exchange 0.086 -> 0.086; Weather 0.199 -> 0.197 | Mixed |
| Add random high-frequency modes to the selected low-frequency modes (ablation) | FEL · Exchange and Weather | MSE: Exchange 0.086 -> 0.087; Weather 0.199 -> 0.196 | Mixed |
| Add multiscale experts (ablation) | FiLM · ETTm2, Electricity, Exchange, Traffic, Weather, and Illness | MSE: Relative: + 0.19%; +0%; 0%; +22%; +0.50%; +22% | Helped |
| Add RevIN normalization (ablation) | FiLM · ETTm2, Electricity, Exchange, Traffic, Weather, and Illness | MSE: Relative: +3.6%; +6.4%; +36%; +48%; +1.5%; +70% | Mixed |
| Train the LPU projection matrix instead of keeping it fixed (ablation) | FiLM · ETTm1 and Electricity | MSE: +4.6%; +2.5%; -2.4%; +24%; +11%; +7.8%; -3.9%; +10% | Mixed |
Closest papers by meaning
- Rethinking Fourier Transform from A Basis Functions Perspective for Long-term Time Series Forecasting (NeurIPS 2024 · similarity 0.84)
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.83)
- Scaling Law for Time Series Forecasting (NeurIPS 2024 · similarity 0.83)
- Timer: Generative Pre-trained Transformers Are Large Time Series Models (ICML 2024 · similarity 0.83)
- Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts (ICLR 2025 · similarity 0.83)
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.