Forecasting SPY using TimeGPT

TimeGPT is a production-ready generative pretrained transformer for time series. It’s capable of accurately predicting various domains such as retail, electricity, finance, and IoT. Let's try forecasting next week's SPY closing prices using TimeGPT and see how it performs.

The data

We will use last year's SPY daily adjusted closing prices for this exercise, a total of 252 data points.

Forecast using TimeGPT

We will try to forecast the next 7 days' SPY closing prices using TimeGPT.

First we need to install nixtla by running pip install nixtla.

Then we initialize the NitxlaClient which requires an API key to access TimeGPT. You can get your API key here.

from nixtla import NixtlaClient

nixtla_client = NixtlaClient(
    # defaults to os.environ.get("NIXTLA_API_KEY")
    api_key = 'YOUR_NIXTLA_API_KEY'
)

Then we can forecast SPY closing prices for the next 7 days.

forecast = nixtla_client.forecast(df, h=7, freq='B', time_col='date', target_col='SPY')

Here are the results:

What about uncertainty?

The above forecast results are point predictions. Usually the reality does not exactly match forecast. If we know the range of the prediction, we can make better informed decisions. We can use multiple confidence levels (50%, 80% and 90%) with TimeGPT.

forecast = nixtla_client.forecast(df, h=7, freq='B', time_col='date', target_col='SPY', level=[50, 80, 90])

Here are the results:

Performance

How do we know how good the predictions are? Let's look at the predictions on historical data.

The trend of predictions largely track the historical prices, and historical prices mostly fall within 50% confidence interval. That's a good sign. In the next post, we will evaluate TimeGPT's performance using more sophisticated cross validation technique.

Summary

In this post, we use TimeGPT to predict the next 7 days' adjusted closing prices of SPY, and qualitatively evaluated its performance - historical prices largely fall within 50% confidence interval, which is promising. Stay tuned for more experiments with TimeGPT.