ts-exogenous-forecast

Time SeriesTime-Series-Library

Description

Exogenous Variable Forecasting: Custom Model Design

Objective

Design and implement a custom deep learning model for time series forecasting with exogenous (external) variables. Uses features=MS: all variables as input, predict only the target (last dimension). Your code goes in the Model class in models/Custom.py. Three reference implementations (DLinear, PatchTST, iTransformer) are provided as read-only.

Evaluation

Trained and evaluated on three datasets with MS features:

ETTh1 (7 → 1, hourly electricity data)
Weather (21 → 1, weather observations)
ECL (321 → 1, electricity consumption)

All use seq_len=96, pred_len=96. Metrics: MSE and MAE on the target variable (lower is better). The framework automatically extracts outputs[:, :, -1:].

Code

Custom.py

EditableRead-only

1import torch
2import torch.nn as nn
3
4
5class Model(nn.Module):
6    """
7    Custom model for exogenous variable forecasting (features=MS).
8
9    Forward signature: forward(x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None)
10    - x_enc: [batch, seq_len, enc_in] — all input variables
11    - x_mark_enc: [batch, seq_len, time_features] — time feature encoding
12    - x_dec: [batch, label_len+pred_len, dec_in] — decoder input
13    - x_mark_dec: [batch, label_len+pred_len, time_features] — decoder time features
14
15    Must return: [batch, pred_len, c_out] for forecasting

Additional context files (read-only):

Time-Series-Library/models/DLinear.py
Time-Series-Library/models/PatchTST.py
Time-Series-Library/models/iTransformer.py
Time-Series-Library/layers/AutoCorrelation.py
Time-Series-Library/layers/Autoformer_EncDec.py
Time-Series-Library/layers/Conv_Blocks.py
Time-Series-Library/layers/Crossformer_EncDec.py
Time-Series-Library/layers/Embed.py
Time-Series-Library/layers/FourierCorrelation.py
Time-Series-Library/layers/SelfAttention_Family.py
Time-Series-Library/layers/StandardNorm.py
Time-Series-Library/layers/Transformer_EncDec.py

Results

Model	Type	mse ETTh1 ↓	mae ETTh1 ↓	mse Weather ↓	mae Weather ↓	mse ECL ↓	mae ECL ↓
dlinear	baseline	0.064	0.188	0.006	0.063	0.387	0.451
itransformer	baseline	0.059	0.187	0.001	0.027	0.301	0.405
patchtst	baseline	0.058	0.183	0.002	0.029	0.318	0.395
timexer	baseline	0.057	0.181	0.001	0.027	0.266	0.368
timexer	baseline	0.057	0.183	0.001	0.027	0.262	0.365
anthropic/claude-opus-4.6	vanilla	0.056	0.179	0.001	0.026	0.418	0.478
deepseek-reasoner	vanilla	0.064	0.192	0.001	0.028	0.320	0.416
google/gemini-3.1-pro-preview	vanilla	0.062	0.189	0.001	0.027	0.339	0.441
openai/gpt-5.4-pro	vanilla	0.056	0.179	0.001	0.025	0.310	0.398
anthropic/claude-opus-4.6	agent	0.065	0.188	0.006	0.063	0.388	0.457
deepseek-reasoner	agent	0.068	0.200	0.001	0.027	0.301	0.406
google/gemini-3.1-pro-preview	agent	0.057	0.182	0.001	0.027	0.279	0.384
openai/gpt-5.4-pro	agent	0.056	0.179	0.001	0.026	0.260	0.364
qwen/qwen3.6-plus:free	agent	0.058	0.183	0.001	0.028	0.497	0.535

Agent Conversations

anthropic/claude-opus-4.6

6 steps

deepseek-reasoner

9 steps

google/gemini-3.1-pro-preview

6 steps

openai/gpt-5.4-pro

5 steps