A user who has already cancelled his subscription is still wondering/frustrated why he's not getting better results. I replied with our general recommendations, but is their anything brief I can also share with him about his particular use case? Maybe @Yibei you have some advice? > I recently created an account and started using TimeGPT for my forecasting tasks. As a quant trader, I have been exploring its capabilities and followed your quickstart guide to input a time series and forecast the next 4 values. Please note that I am forecasting a range-bound cyclic time series, not a core stock price. > However, to my surprise, the results I received are worse than those from simple Linear AutoRegression, and far below the performance of Chronos T5 Large. I suspect I might have made an error somewhere, as these results do not seem to justify the high cost of the subscription. I would appreciate clarification on the following points: > 1. I am using a daily time series with a total of 8,000 points (SPX close-extracted series). Is it mandatory to provide a date column, or can the row serial number suffice? > 2. I can provide the dates, which is simple, but how does the model estimate future trade dates when forecasting (because dates are not uniform because of the weekends and holidays)? Is there a feature that allows us to input the forecast horizon dates manually (though this doesn't seem an ideal solution)? If the dates are not forecasted correctly, it could significantly impact the results. > 3. Although your platform indicates that the model is ready for forecasting, here are the results I am getting via the direct client forecast method: (I have replace the serial numbers with continuous dates to check the results) > As you can see, the R² is highly negative, while Chronos T5 provides me with an 80-85% R², and Linear AutoRegression (with my manual cyclic features) yields around 88% R². > Could you please guide me on whether I might be doing something wrong in the forecasting process? Additionally, is there a workaround for the date column, given that future dates will have missing values (e.g., weekends) as I only plan to add the serial numbers (which I think are more than sufficient to capture the cyclicity)? I plan to apply this forecasting to 1,000 symbols, and I would consider subscribing to the TimeGPT plan if it offers better forecasts than my current systems. Images of results in thread.

Hi Tracy! I drafted the reply.

Hi! Thanks for reaching out, and we’re happy to assist with your question.

Firstly, the negative R² you’re seeing likely indicates that TimeGPT isn’t fully capturing the patterns in your data. Since you’re working with cyclic time series, I recommend trying the

finetune_steps

parameter in your forecast. This might help TimeGPT align better with the cyclic behavior present in your data. Also, you can decrease the forecast

horizon

to see if the performance improve.

Also, TimeGPT generally performs best when you include a date column. This is important for understanding the temporal structure, particularly with financial data, where trading days can be irregular due to weekends and holidays. To account for the gaps caused by holidays and weekends, try setting

freq = "B"

(business day frequency) in your forecast configuration.

If the issue persists, feel free to share some dummy data with us, and we’d be happy to assist further and calibrate the model to better suit your needs.

Thanks so much. I'll use this. I just won't include the part about better calibrating the model, so he doesn't think we'll do that right away. 😅

Oops sorry, I mean we can look look into his data and try finetune etc.😁

Ah, got it, thanks!