Codesota · Benchmark · M4 CompetitionHome/Leaderboards/Structured Data & Forecasting/Time-Series Forecasting/M4 Competition
Unknown

M4 Competition.

100,000 time series from diverse domains (finance, demo, macro, micro, industry, other).

Paper ↗Leaderboard ↓
§ 01 · SOTA history

Year over year.

§ 02 · Leaderboard

Results by metric.

Found a wrong score or missing run?
Use row edits to send a sourced correction into moderation.
Add / edit result ↗Report issue ↗

smapi

Smapi is the reported evaluation metric for M4 Competition. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Higher is better

Trust tiers for smapiverifiedpapervendorcommunityunverified

Muted rows were not state of the art when published — an earlier or same-year result already scored better.

RankModelTrustScoreYearLinksFix
01TiDE
Long-term time series forecasting with MLP-Mixer encoder-decoder. Results from TimeMixer++ comparison table.
verified13.952023Source ↗Looks wrong?
02DLinearverified13.6392022Paper ↗Code ↗Source ↗Looks wrong?
03PatchTSTverified13.1522022Paper ↗Code ↗Source ↗Looks wrong?
04Autoformer
NeurIPS 2021. Auto-Correlation mechanism for time series decomposition. Results from TimeMixer++ comparison table.
verified12.9092021Source ↗Looks wrong?
05FEDformerverified12.842022Paper ↗Source ↗Looks wrong?
06iTransformer
ICLR 2024. Inverted transformer with attention applied across variables. Results from TimeMixer++ comparison table.
verified12.6842024Source ↗Looks wrong?
07N-HiTSverified11.9272022Paper ↗Code ↗Source ↗Looks wrong?
08LMS-AutoTSF
Dec 2024. Learnable Multi-Scale Decomposition and Integrated Autocorrelation. Average across Yearly/Quarterly/Monthly/Others.
verified11.8712024Source ↗Looks wrong?
09N-BEATS
ICLR 2020. Neural basis expansion analysis. Results from TimeMixer++ comparison table (no ensemble).
verified11.8512020Source ↗Looks wrong?
10TimesNet
ICLR 2023. Transforms 1D time series into 2D space for temporal variation modeling.
verified11.8292023Source ↗Looks wrong?
11TimeMixerverified11.7232024Paper ↗Code ↗Source ↗Looks wrong?
12TimeMixer++verified11.4482024Paper ↗Code ↗Looks wrong?
13ES-RNNverified11.3742020Paper ↗Code ↗Looks wrong?

mase

Mase is the reported evaluation metric for M4 Competition. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Higher is better

Trust tiers for maseverifiedpapervendorcommunityunverified

Muted rows were not state of the art when published — an earlier or same-year result already scored better.

RankModelTrustScoreYearLinksFix
01DLinearverified2.102022Paper ↗Code ↗Source ↗Looks wrong?
02PatchTSTverified1.952022Paper ↗Code ↗Source ↗Looks wrong?
03TiDE
Long-term time series forecasting with MLP-Mixer encoder-decoder. Results from TimeMixer++ comparison table.
verified1.942023Source ↗Looks wrong?
04Autoformer
NeurIPS 2021. Auto-Correlation mechanism for time series decomposition. Results from TimeMixer++ comparison table.
verified1.772021Source ↗Looks wrong?
05iTransformerverified1.762023Paper ↗Code ↗Source ↗Looks wrong?
06FEDformerverified1.702022Paper ↗Source ↗Looks wrong?
07N-HiTS
AAAI 2023. Neural hierarchical interpolation for time series. Results from TimeMixer++ comparison table.
verified1.612022Source ↗Looks wrong?
08LMS-AutoTSF
Dec 2024. Learnable Multi-Scale Decomposition and Integrated Autocorrelation. Average across Yearly/Quarterly/Monthly/Others.
verified1.592024Source ↗Looks wrong?
09TimesNet
ICLR 2023. Transforms 1D time series into 2D space for temporal variation modeling.
verified1.582023Source ↗Looks wrong?
10N-BEATSverified1.562019Paper ↗Code ↗Source ↗Looks wrong?
11TimeMixer
ICLR 2024. Decomposable multiscale mixing for time series forecasting.
verified1.562024Source ↗Looks wrong?
12ES-RNNverified1.542020Paper ↗Code ↗Looks wrong?
13TimeMixer++verified1.492024Paper ↗Code ↗Looks wrong?

Owa

Owa is the reported evaluation metric for M4 Competition. Codesota tracks published model scores on this metric so readers can compare state-of-the-art results across sources and model families.

Higher is better

Trust tiers for Owaverifiedpapervendorcommunityunverified

Muted rows were not state of the art when published — an earlier or same-year result already scored better.

RankModelTrustScoreYearLinksFix
01DLinearverified1.052022Paper ↗Code ↗Source ↗Looks wrong?
02TiDE
Long-term time series forecasting with MLP-Mixer encoder-decoder. Results from TimeMixer++ comparison table.
verified1.022023Source ↗Looks wrong?
03PatchTST
ICLR 2023. Transformer with patch tokenization. Results from TimeMixer++ comparison table.
verified1.002023Source ↗Looks wrong?
04Autoformer
NeurIPS 2021. Auto-Correlation mechanism for time series decomposition. Results from TimeMixer++ comparison table.
verified0.942021Source ↗Looks wrong?
05iTransformerverified0.932023Paper ↗Code ↗Source ↗Looks wrong?
06FEDformer
ICML 2022. Frequency enhanced decomposed transformer. Results from TimeMixer++ comparison table.
verified0.922022Source ↗Looks wrong?
07N-HiTSverified0.862022Paper ↗Code ↗Source ↗Looks wrong?
08N-BEATSverified0.852019Paper ↗Code ↗Source ↗Looks wrong?
09LMS-AutoTSF
Dec 2024. Learnable Multi-Scale Decomposition and Integrated Autocorrelation. Average across Yearly/Quarterly/Monthly/Others.
verified0.852024Source ↗Looks wrong?
10TimesNetverified0.852022Paper ↗Code ↗Source ↗Looks wrong?
11TimeMixerverified0.842024Paper ↗Code ↗Source ↗Looks wrong?
12ES-RNN
M4 competition winner (2018). Hybrid exponential smoothing + RNN by Slawek Smyl (Uber). OWA=0.821.
verified0.822018Source ↗Looks wrong?
13TimeMixer++
ICLR 2025. Multi-resolution time imaging + decomposable multiscale mixing. Achieves OWA 0.821 matching original M4 competition winner.
verified0.822024Source ↗Looks wrong?
§ 04 · Submit a result

Add to the leaderboard.

← Back to Time-Series Forecasting