Sanity check

< supervised long-term forecasting results of my base model* >

 

* base model: GPT-2 without injecting any additional information

 

The backbone model can be any LLM, but I used GPT-2 with 6 layers as default for simplicity. 

 

I may conduct an ablation study on different LLM model variants and sizes. Several previous studies have demonstrated that the scaling law also applies to time-series forecasting in relation to the number of model parameters and the size of the training corpus.

 

content length 512 / forecasting horizon 96 

---

1) ETTh1 : training epochs 10

512_96_MyModel_ETTh1_sl512_pl96_dm32_nh8_df128_0 
test on the ETTh1 dataset: mse: 0.3996824, mae: 0.4219979

 

2) ETTm1: training epochs 10

512_96_MyModel_ETTm1_sl512_pl96_dm32_nh8_df128_0 
test on the ETTm1 dataset: mse: 0.3175505, mae: 0.3626745

 

3) Weather : training epochs 1 

512_96_MyModel_Weather_sl512_pl96_dm32_nh8_df32_0 
test on the weather dataset: mse: 0.1589350, mae: 0.2111652

 

4) Electricity: training epochs 1

512_96_MyModel_ ECL _sl512_pl96_dm32_nh8_df32_0 
test on the electricity dataset: mse: 0.1420454 , mae: 0.2483649

---

Some visualization (cherry picking)

1) ETTh1

2) ETTm1

3) Weather