There are several advantages/disadvantages of Random Forest or comparisons with ARIMA. I am looking for advices regarding time series forecasting with Random Forest. My current results show that Random Forest has some problems with time series.
I am not familiar with applying random forest to Time series forcasting, rather I have idée reccurent neural network RNN ans particurlarly LSTM that shows very interesting results
Random Forest, as well as other tree-based methods, generally show poor performance on time-series with trends. The reason is that those algorithms are unable to extrapolate values beyond the min/max values that they've observed in the training data. A potential modification that could solve this issue is detrending of time-series before applying Random Forest. For example, you could fit a simple linear regression to the original series vs time and subtract the resulting line. The forecast then is a summation of a simple linear trend + Random Forest predictions.
You can read more on applying Machine Learning to time series forecasting in my article - https://medium.com/open-machine-learning-course/open-machine-learning-course-topic-9-time-series-analysis-in-python-a270cb05e0b3
Method of Random forest is suited for classification problems. For time-series forecasts, mainly when a large number of variables are possibly affecting values, any benchmark technique must be validated for performance. Literature may have instances of mixed results. Read more:
Article Comparison of ARIMA and Random Forest time series models for...
Thank you...but i do not agree that RF is only suited
for classification problems.
Since you will mostly look to use bagging while employing random forest, the sequential order of input features may be lost. For reasons pertaining to generalization, it may not be a good idea to add time stamp as part of your feature set. Hence it is necessary to detrend or make sure the time series is covariate stationary using simple transformations .e.g differencing. You should also look to add time-lagged versions of relevant features. Finally, while Random forest/Ensemble based algorithms are extremely powerful, they are not the best for time series problems since they make no assumptions regarding distribution of data and also while splitting a node, averaged values of all samples are considered thus ignoring any time varying trends. Hence they cannot be used efficiently for time series predicitions very well( extrapolation in time) since there may be values in test/validation set which lie outside range of training samples. You may further look into RNNs e.g. LSTM for such problem statements
- Badges
- Science topic
- Similar topics
- Control Systems Engineering
- Filtering