Why TimeMixer?

TimeMixer is a fully MLP-based architecture that achieves state-of-the-art performance on 18 time series benchmarks without attention, recurrence, or convolution stacks. It works by:

1. Decomposing time series into seasonal and trend components at multiple temporal scales
2. Mixing seasonal patterns bottom-up (fine → coarse) and trend patterns top-down (coarse → fine)
3. Predicting by ensembling complementary forecasts from each scale

Key result: Outperforms PatchTST, TimesNet, iTransformer, and DLinear while using fewer parameters and less GPU memory.

Past-Decomposable-Mixing (PDM) + Future-Multipredictor-Mixing (FMM)

Installation

pip install timemixer-tf

For GPU support, install TensorFlow with CUDA:

pip install tensorflow[and-cuda] timemixer-tf

Quick Start

5-Minute Example

from timemixer_tf import TimeMixerConfig, TimeMixer
import numpy as np

# Configuration matching the paper's ETT benchmark
config = TimeMixerConfig(
    task_name="long_term_forecast",
    seq_len=96,          # Look-back window
    pred_len=96,         # Forecast horizon
    enc_in=7,            # Number of input features
    c_out=7,             # Number of output features
    d_model=16,          # Model dimension
    e_layers=2,          # PDM blocks
    down_sampling_layers=3,
    down_sampling_window=2,
)

model = TimeMixer(config)

# [batch, seq_len, features] → [batch, pred_len, features]
x = np.random.randn(32, 96, 7).astype(np.float32)
x_mark = np.zeros((32, 96, 4), dtype=np.float32)  # time features

prediction = model(x, x_mark, training=False)
print(prediction.shape)  # (32, 96, 7)

Supported Tasks

Channel Independence

Set channel_independence=1 (default) to treat each feature independently — recommended for most datasets. Set channel_independence=0 for cross-channel mixing with fewer features.

config = TimeMixerConfig(channel_independence=0)  # cross-channel
config = TimeMixerConfig(channel_independence=1)  # independent (default)

Architecture

Input [B, T, C]
    │
    ├─ Multi-scale down-sampling ──► [T, T/2, T/4, T/8, ...]
    │
    ├─ RevIN normalization (per scale)
    │
    ├─ Data Embedding (conv1d + time features)
    │
    ▼
┌─────────────────────────────────────────┐
│  Past Decomposable Mixing (PDM) × L     │
│                                         │
│  For each scale:                        │
│    ├─ Decompose → season + trend        │
│    ├─ Season: bottom-up mixing          │
│    │    (fine → coarse aggregation)      │
│    └─ Trend: top-down mixing            │
│         (coarse → fine refinement)       │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│  Future Multipredictor Mixing (FMM)     │
│                                         │
│  Predict from each scale → ensemble     │
└─────────────────────────────────────────┘
    │
    ▼
Output [B, pred_len, C]

Training on ETT Benchmarks

import pandas as pd
from sklearn.preprocessing import StandardScaler
import tensorflow as tf

# 1. Load data
df = pd.read_csv("ETTh1.csv")
data = StandardScaler().fit_transform(df.values[:, 1:])

# 2. Create sequences
seq_len, pred_len = 96, 96
xs = np.lib.stride_tricks.sliding_window_view(data, seq_len, axis=0)
xs = np.swapaxes(xs[:n], 1, 2).astype(np.float32)

# 3. Build model
model = TimeMixer(TimeMixerConfig(
    task_name="long_term_forecast",
    seq_len=seq_len, pred_len=pred_len,
    enc_in=7, c_out=7,
))

# 4. Train with Keras
model.compile(optimizer="adam", loss="mse")
model.fit(train_dataset, epochs=10)

Results

Verified against the official PyTorch implementation on ETT benchmarks:

Paper reference: ETTh1 avg MSE 0.447 (ICLR 2024 Table 2)

Differences from PyTorch Version

All differences are implementation-level; the mathematics is identical.

Citation

If you use this implementation in your research, please cite the original paper:

@inproceedings{wang2023timemixer,
  title={TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting},
  author={Wang, Shiyu and Wu, Haixu and Shi, Xiaoming and Hu, Tengge and
          Luo, Huakun and Ma, Lintao and Zhang, James Y and ZHOU, JUN},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2024}
}

License

Apache 2.0 — see LICENSE.

Original PyTorch implementation: kwuking/TimeMixer (MIT licensed).