Time Series Prediction (Forecast into the Future)
timothy_rij
Member Posts: 3 
Learner II
Hi All,
I have spent the past few days learning the basics of RapidMiner and looking through the tutorials and existing posts. There are a lot of examples on time series prediction comparing existing data against predicted/forecasted data and scoring/ranking the two streams against each other.
I am yet to find any discussion on “predicting into the future” where there is no available data. Eg use only the past few years of daily data to predict the next 100 days. As always, my question relates to the stock price of a single company.
I have setup my design view to use an ANN as per below.
However I would like RapidMiner to predict the next 100 days etc into the future where I have no data as shown in the blue drawn line below. Has anyone done this before or would know of an example I can examine?
Answers
Time series forecasting processes are done differently from the usual classification or regression methods.
There's an entire course in the Academy about forecasting:
https://academy.rapidminer.com/courses/time-series-analytics
Please check this out.
Your process does a windowing for 7 days. This builds a model that can only predict the next day's value given the last 7 values. This could be extended in theory to learn a model from the last 200 days, with days -200 to -101 being the input attributes and days -100 to -1 the labels. These models could be applied to current data in a rolling manner. But this is cumbersome and you won't get good results from these models in the stock market setting, as stock prices depend on more factors than just the past values.
Regards,
Balázs
I do agree with you in that "stock prices depend on more factors than just the past values", hence my data set includes 21 fields with opening price being one field and the remaining 20 fields being other related data.
I've seen presentations at conferences. People have been testing and comparing methods for predicting stock prices for decades. The results were underwhelming.
In this field you're up against Wall Street companies with large rooms full of Harvard PhD holders. And even those companies don't have the magic model to beat the market.
You can do a windowing on all 20 attributes and build your models with the windowed, "old enough" data. But don't expect a prediction horizon of 100 days with something as volatile as stock prices to be reliable.
Regards,
Balázs