Hi Folks, I'm experimenting with TimeGPT model, and I’m trying to forecast a table that has multiple unique_ids, which might depend on each other (like historical exogenous values). This is a sample data I’m working with:

Month, Currency Pair, Account, Activity
2022-01-01, EURUSD, 1000 (Cash), 3000
2022-01-01, EURUSD, 2000 (Payables), -1000
2022-01-01, EURUSD, 3000 (Receivables), 5000
2022-02-01, EURUSD, 1000 (Cash), 4900
2022-02-01, EURUSD, 2000 (Payables), -3000
2022-02-01, EURUSD, 3000 (Receivables), 2000
2022-03-01, EURUSD, 1000 (Cash), -2000
2022-03-01, EURUSD, 2000 (Payables), -10000
2022-03-01, EURUSD, 3000 (Receivables), 19000
2022-01-01, USDCAD, 1000 (Cash), 4500
2022-01-01, USDCAD, 2000 (Payables), -1500
2022-01-01, USDCAD, 3000 (Receivables), 7500
2022-02-01, USDCAD, 1000 (Cash), 7350
2022-02-01, USDCAD, 2000 (Payables), -4500
2022-02-01, USDCAD, 3000 (Receivables), 3000
2022-03-01, USDCAD, 1000 (Cash), -3000
2022-03-01, USDCAD, 2000 (Payables), -15000
2022-03-01, USDCAD, 3000 (Receivables), 28500
2022-01-01, USDINR, 1000 (Cash), -1500
2022-01-01, USDINR, 2000 (Payables), 500
2022-01-01, USDINR, 3000 (Receivables), -2500
2022-02-01, USDINR, 1000 (Cash), -2450
2022-02-01, USDINR, 2000 (Payables), 1500
2022-02-01, USDINR, 3000 (Receivables), -1000
2022-03-01, USDINR, 1000 (Cash), 1000
2022-03-01, USDINR, 2000 (Payables), 5000
2022-03-01, USDINR, 3000 (Receivables), -9500

The ds column is Month, and the y column is Activity. The main unique_id column is Currency Pair, but I also have the Account column that distinguishes the Activity. To use the TimeGPT APIs, I tried to pivot the table on Account, and have something like this:

Month, Currency Pair, 1000 (Cash), 2000 (Payables), 3000 (Receivables)
2022-01-01, EURUSD, 3000, -1000, 5000
2022-01-01, USDCAD, 4500, -1500, 7500
2022-01-01, USDINR, -1500, 500, -2500
2022-02-01, EURUSD, 4900, -3000, 2000
2022-02-01, USDCAD, 7350, -4500, 3000
2022-02-01, USDINR, -2450, 1500, -1000
2022-03-01, EURUSD, -2000, -10000, 19000
2022-03-01, USDCAD, -3000, -15000, 28500
2022-03-01, USDINR, 1000, 5000, -9500

I tried to forecast the “1000 (Cash)” column:

forecast = nixtla_client.forecast(df=df, time_col='Month', id_col='Currency Pair', target_col='1000 (Cash)', h=3)

but it doesn’t take in consideration the other two “_2000 (Payables)”_, “_3000 (Receivables)_” for calculating the forecast, nor it adds them in the output:

unique_id, ds, TimeGPT
EURUSD, 2022-04-01, -1397.99
EURUSD, 2022-05-01, -345.228
EURUSD, 2022-06-01, 1201.812
USDCAD, 2022-04-01, -2096.98
USDCAD, 2022-05-01, -517.842
USDCAD, 2022-06-01, 1802.718
USDINR, 2022-04-01, -165.96
USDINR, 2022-05-01, -698.591
USDINR, 2022-06-01, -869.486

What can you recommend for my scenario where the value to forecast could depend on N other columns (that I have in the history), and the model automatically infers the relationships/weights, forecasting them all at once? I see in the TimeGPT forecast function definition, the

id_col

and

target_col

are a single string, and I couldn't find any way to provide multiple ones. I see there is a way to add exogenous variables, but they all need to be future values, while I need to forecast them.

Hi Luca! So, my understanding is that you actually want to forecast each account type for each currency pair. Therefore, I would suggest that you build a series for each currency pair and account type. This means that your id column would be something like: USDCAD-1000 USDCAD-2000 USDCAD-3000, etc. and you do that for each currency pair and accoun type. I think it will also give you better results, since payables seems to contain mostly negative values. That way, TimeGPT will forecast each currency pair and account type at once. Let me know if it makes sense and if that helps you! 🙂

It kinda does (and that's what the forecast of the pivoted table returns), but the reality is that each account value (the activity) may be interdependent of the values in the other accounts. For example,

Cash = Receivables - Payables +- DELTA

I would like the model to infer that considering all the values at the same time, not just the individual series. Also the numbers in this sample a just randomly generated, they do not represent the real data.

When you pass multiple series, it already does that. It will use the information from all series to make predictions.

but how do I pass multiple series? In the TimeGPT forecast function definition, the

id_col

and

target_col

are a single string, and I couldn't find any way to provide multiple ones. Should I merge

Currency Pair

with

Account

column, and pass that as

id_col

?

This is done for your. It checks for unique values in the id column. If there are many unique values, it knows that multiple series were passed, and it forecasts all of them. If there is a single unique value in the id column, then one series is predicted.

Got it! I merged the two columns in a single

unique_id

column, and the forecast outputs everything as expected, and you are saying that the prediction will take into consideration the potential inter-relationships / weights between the various series. That's awesome! I have another question: I see in the Azure AI page that the possibility to deploy your model in a private Azure environment is coming soon. We would like to use TimeGPT to forecast private company data, and we need a private environment before we have permission to try anything. How soon can we expect Microsoft to offer your model in Azure AI Studio?

I will let @Cristian (Nixtla) answer that question 😅, I'm not aware of all those details. I am glad I could help and looking forward to have you as a TimeGPT user 🙌

Thanks Marco! TimeGPT can forecast multiple time series of a dataset, but each forecast is done independently, so it does not leverage potential inter-relationships. The approach you should follow is to add them as exogenous variables, as you were initially doing. However, you need to pass the

X_df

dataframe for TimeGPT to consider them. The model only currently supports exogenous variables for which you have data on the forecasting horizon (future variables). You can incorporate them in two stages. First, you forecast the exogenous covariates with TimeGPT. Second, pass the forecasts to the

X_df

parameter of the forecast method. We are working on adding historic variables as well.

The model only currently supports exogenous variables for which you have data on the forecasting horizon (future variables)

Ouch, all the series may be interdependent... I would be required to arbitrary choose which ones may not be dependent, then forecast them, and use that forecast as exogeneous covariates with

x_df

parameter, which is not ideal for my scenario. I see that your NeuralForecast supports *historic* exogenous variables, is that something you would consider adding to TimeGPT? Also, could you please tell me about Azure AI?

Yes, it will support historical exogenous covariates in the future

That is awesome! Please make the api as easy as providing the column name of the exogenous covariates, instead of passing a separate dataframe. Ps: can you please tell me about the availability of a private implementation in Azure AI?