lookback len choosing in finetunomg with different frequencys

[Zenki229]

First, Kronos is truly astonishing and interesting—thank you for releasing it as open source!

To finetune with different freqeuncies, could we refer table 4 in data preprocessing sec? That is to say, we can select different max_model_len for different freqs?

[shiyu-coder]

Thanks! Of course, you can actually mix any lookback length for training, even exceeding 512 (but may require a larger dataset for finetuning).

[Zenki229]

Thank you! I wonder if you try useing LoRA to finetune large model with limited number of datasets. I am trying to compare performance finetuning the model(base size, LoRA) with the model(tiny or small size, full).

[zhenqiao0277]

Thanks! Of course, you can actually mix any lookback length for training, even exceeding 512 (but may require a larger dataset for finetuning).

Thank you for sharing the code. How do you suggest setting the lookback window for BTC/USDT with a 5-minute candle size and a 30-minute predict length?

[shiyu-coder]

Thank you! I wonder if you try useing LoRA to finetune large model with limited number of datasets. I am trying to compare performance finetuning the model(base size, LoRA) with the model(tiny or small size, full).

Sorry, I have not tested the effectiveness of LoRA finetuning. Since the model size is not large compared with llms.

[shiyu-coder]

Thanks! Of course, you can actually mix any lookback length for training, even exceeding 512 (but may require a larger dataset for finetuning).

Thank you for sharing the code. How do you suggest setting the lookback window for BTC/USDT with a 5-minute candle size and a 30-minute predict length?

The specific value is not easy to say. It is recommended to test the optimal lookback window value based on the actual situation

[zhenqiao0277]

@shiyu-coder
Thanks for the reply. Additionally, if the predictor shows obvious overfitting during training, how would you suggest adjusting the training parameters?