Not sure what to say to this one.

Hello,When using your company's model TimeGPT, I found a problem. I found that the predicted value leads the real value. The first picture is the result of Bitcoin prediction on your company's timegpt official website , and the second is the result of my prediction .These both have obvious predictions leading the real value. Is this normal? Or does timegpt have this amazing performance?

My understanding is: 1. TimeGPT is doing a great job🥳, it's performing better than naive model 2. Historical forecast uses a rolling window to do forecast. It makes forecast one step ahead everytime and concat the result. That's why it's not deviate from real value over long term. Am I right?

Funny, because I have been working on that tutorial today. Tbh, I don't think you can predict something like Bitcoin, not with TimeGPT and certainly not with ARIMA. We added that tutorial because some users believe you can make those predictions.

Thanks @Yibei and @Mariana Menchero, great explanation!

No problem Tracy. I recommend you explain to the user what Yibei just mentioned and gently suggest him that he should benchmark against a random walk 😅

I wonder if this is any information we should include in the bitcoin tutorial if it's not there already?

In the second to last section we have a paragraph that says that "_predicting the future prices of financial assets is a challenging task, especially for assets like Bitcoin. However, for those who need or want to forecast these assets,

TimeGPT

can be a powerful tool that simplifies the forecasting process._" Do you think we should be more explicit?

A sentence that combines the information from you and Yibei could be good. Something like predicting the future prices of financial assets is a challenging task, especially for assets like Bitcoin. The predictions in this tutorial are very accurate, because we're doing historical forecasting. The real challenge is forecasting the price of Bitcoin for the upcoming days, not its historical price. For those who need or want to try to forecast these assets,

TimeGPT

can be an option that simplifies the forecasting process.

Great, let me add this. This question couldn't have come at a better time because I had to change the way data is downloaded for that tutorial, so I have an open PR for it 😅

I dont think that the fact they are historical forecasts explains the good performance, right? The historical forecasts are only seeing their corresponding past values so it shouldn't be easier. TimeGPT was not trained on this data and we always do historical forecasts for cross validation.

I see. What do you think of this new version @Cristian Challu? Predicting the future prices of financial assets is a challenging task, especially for assets like Bitcoin. The predictions in this tutorial may appear accurate, mainly because they align with recent historical data and the model updates with new values at short intervals, avoiding significant deviations. However, the true challenge lies in forecasting Bitcoin's price for the upcoming days, not just its historical performance. For those who need or want to try to forecast these assets, TimeGPT can be an option that simplifies the forecasting process.

looks good