Hi all, I've been experimenting with timegpt api for a while. it is pretty straight forward to use the api and am able to generate predicting on my asset price data. I have ml experience with non-time series data , but there's are special things about time series data forecasting with timegpt I'm catching up and would like to understand : 1. Do I need to worry about choosing the best lookback period before giving the data to timegpt.forecast? or i just feed all the historical data i have (eg. 3 year daily closing price). If I need to decide a best lookback period, instead of trying random numbers and test on timegpt.forcast, what is the best approach to decide which lookack period to choose ? should I look at the historical forecast charts and performance metrics of each lookback period-interval combination first, and use some techniques to decompose the trend , seasonality etc of my data first. and then give best guess of ookback period to timegpt.forecast? Do you have any suggestions on finding a good lookback period to feed to timegpt? 2. I run grid search on a few parameters and measure a few metrics like MAE, MSE, MAPE, DA, for example gridsearch over :

lookback_period=[60,90,365]
finetune_steps =[10,50,100]
pred_horizon=[7,14,30]
freq='D'

should i grid search all other paramters to find the best combination first, and then adjust the finetune_steps . Or is it ok to do the all at once like above? 3. when i include exogeuous variables, should I check their multi-collinearity and cross-correlations (eg. using emsemble method based on random decision forest) with the target and only select the ones with low cross-correaltions and no multi-colinearity first, and give them to timegpt? I see timegpt has apis to display weights of each exogeuous variables. but i'm not sure if I am intending to try 100 + exogeuous variable, should I do a pre-screening first ? and then use the timegpt's weighting results to further adjust and removal ? Many thanks !