I have one year of wind speed data, collected at hourly intervals. I am trying to train a forecasting model using this data. I want to be able to generate rolling forecasts up to 7 days in the future. How do I decide the amount of data to use in my training set. Should I use the whole one year data or should I use a subset of it. How can one determine this amount of data to be used while training. I understand that having more data points than the seasonality of the data is always good in time series forecasting. But the data is rather noisy and no seasonality is observed other than diurnal seasonality.
Hi Bhavya,
Based my experience and mentorship from highly experienced data science experts in machine learning, the common practice splits the dataset into two sets in the following proportions - 75% training set (to optimise the propagations, and 25 test set used for prediction.
With time series data, you have to be careful when splitting data randomly since you need to presever the sequential order within test and training sets. Those papers might help you out:
Article On the use of cross-validation for time series predictor evaluation
Preprint Evaluating time series forecasting models: An empirical stud...
I agree with Kai Heinrich about splitting data randomly. Cross-validating the time-series model is cross-validation on a rolling basis. Start with a small subset of data for training purpose, forecast for the later data points and then checking the accuracy for the forecasted data points.
The procedure would be something like this:
- fold 1 : training [1], test [2]
- fold 2 : training [1 2], test [3]
- fold 3 : training [1 2 3], test [4]
- fold 4 : training [1 2 3 4], test [5]
- fold 5 : training [1 2 3 4 5], test [6]
For detailed information:
https://medium.com/@soumyachess1496/cross-validation-in-time-series-566ae4981ce4
Normally in Machine learning we go for 70-30 or 80-20 train test data split and if you want to validate your model then 60-20-20 split. For time series data you need to check if the data stays stationary or not because of splitting. You can use holdout and k - fold cross validation method for splitting dataset. If there is only diurnal seasonality present then I would suggest that you split your data based on day and hours and compare your model accuracy.
I don't agree fully with most of the previous answers, except the last one. Given the seasonal behaviour of your data, ideally, you should have more than one year in the training set and use the remaining data for validation. Try having longer time series. If it is not possible, you can subdivide your dataset in two parts or more, say, and treat them separately, considering the first weeks for training and the remaining ones for validation. I have only used wind speed data once so I am not an expert on this.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series