Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hello! Someone asks

> I would like to fine-tune TimeGPT on a database 1 and then test it on a database 2 it has never seen, with the memory of the fine-tuning done earlier. Would that be possible?
Draft response in thread. I already sent them the link to the <https://docs.nixtla.io/docs/tutorials-fine_tuning|fine-tuning tutorial>.

Yes, that is possible. You would create a new model with the fine tuning, e.g.

```timegpt_fcst_finetune_df = nixtla_client.forecast(
    df=df, h=12, finetune_steps=10,
    time_col='timestamp', target_col='value',
)```
then you would use that model with your database2, 'df2'

```nixtla_client.plot(
    df2, timegpt_fcst_finetune_df, 
    time_col='timestamp', target_col='value',
)```

I don't think we can do that unless df1 and df2 are same time series splited into train and test set. In that case, we can simply finetune on df1 and make forecast for timestamps in d2, adding exogenous variables in d2 as `future_ex_vars_df` , then comparing the forecast result with target values in df2.