This site uses different types of cookies, including analytics and functional cookies (its own and from other sites). To change your cookie settings or find out more, click here. If you continue browsing our website, you accept these cookies.
I want to test an ARIMA / covariate model (one-day period) that can be used in the future, every day, to predict hospital bed requirements several days ahead. I have five years of daily data, using four years to build the model, and and the last year as the holdout sample to test it. The model comparison tool is the way to do that, but I believe it is only doing a one-period ahead prediction to compare with the actual data. Is that correct? I'd like to use the model to predict, say, requirements 4 days from now and test that against the holdout data. Is there any way to do that?
I'm not a TS expert by any stretch, but from the description of the Model Comparison Tool, it looks as though it works for classification and regression models, but not for Time Series.
This, for me, seems to make sense... if you build a time series based on years 1,2,3,4, and then compare actuals to predictions for year 5... but then still use the same model for the additional days into year 6, that's not quite what Time Series is all about, I don't think, because you would definitely want to use the actual data from year 5 in the model predicting year 6. Unfortunately I'm not truly an expert... I do think that if you generate numerous models based on "using x days in the past to predict y days into the future," and you do this on a rolling scale over your past data, then you can look for what the best value of "x" is (and various other hyperparameters in the model)... that might be an approach worth considering, for determining the best predictive model for the need you described.
Hi @Rkrider, have you looked at the TS Covariate Forecast tool? It provides forecasts from an ARIMA model that uses covariates. The number of periods to forecast is determined by the number of periods of covariate data provided.