Any practical (e.g. easily implementable in R) methods for predicting the values for a given time series that account for additional/external variables influence ?
For example, the no. of orders for a restaurant business can be treated as time-series - but the future orders might get affected by any holidays, discount offers and weather so on. Given data about the no. of orders for each day and the weather on that day for the past one year, how to predict the no. of orders for the next couple of days given the weather forecast?
- Data Input: {Day-wise Order Count, Weather}
- Prediction Output: Given weather forecast, {Day-wise Predicted Order Count}
Similarly, load on a power grid can be viewed as a time-series and predicted based on history, but often any upcoming festivals and special occasions may influence the usage.
Any methodology to take these additional variables into consideration while predicting time-series?
Try Using Wilcoxon signed rank test for outlier analysis and fast convergence together with particle swarm optimization or intelligence technique for learning the model. but before proceeding try to find the independent and depended attribute for prediction. For analysing you time series data, you can also use Moving Average.
Check this paper, i refered the same
http://www.eecs.umich.edu/db/files/sigmod07timeseries.pdf
The model of continuous removal of outlier results in better prediction and convergence
Dear Mr Palem, in my humble opinion there is no "best method". I guess the best method is the best suited to the problem domain you are trying... I suggest you take a look at Weka Times Series Forecasting (http://wiki.pentaho.com/display/DATAMINING/Time+Series+Analysis+and+Forecasting+with+Weka) and Overlay data section. Perhaps it may be useful to you.
Best regards.
One approach is the Bayesian statistics that have been able to integrate external information. But it is difficult to implement without the statistical assistance of a statistician.
I have't found conditional prediction of a time series with external factors in R. The xreg argument is accepted but not used for prediction. The conditional prediction can be implemented as a three-step process, including independent prediction and linear correlation approximation. The first step is to construct independent prediction data. For each date, we calculate a prediction value based on the real historical data. The second step is to create linear model, we will calculate the real data as a function of the predicted data and external factors. Finally, we run the linear model on the predicted value (real data are unknown) and external factors (know.) The result depends on the historical power consumption data and external data.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining