0

My training data contains the prices of the stock and 40 masked features. These masked features are also present in my test data. I want to predict the price column in the test data. Can I solve it as a normal supervised learning problem and not treat it as a time series problem as I have enough independent variables to predict the target variable in test data.

More importantly, how should I approach this problem.

boredaf
  • 98
  • 1
  • 8
  • 1
    Your data is not in time series format (it is not constantly distributed ) so you can not do forecasting on it. – Zeeshan Jan 03 '19 at 06:26
  • Assuming your data is formatted correctly, then based on the plot you shared, your data doesn't have any regular structure (e.g. seasonality, trend, etc...) and so no forecasting method is going to give you good results. Please clarify if you have the correct format or not? You say 1500 samples, at what frequency? – Alex Kinman Jan 03 '19 at 18:46

2 Answers2

1

Also, I doubt that the data which I have can be used for time series forecasting or not.

I am attaching a screenshot of vibration scatter plot.

Data looks strange (>1 value for some values on the x-axis). Why scatter?

I have daily vibration data of compressor(around 1500 samples) and want to forecast it for another 30 days. I tried ARIMA but it is giving poor results.

Actually, there are several approaches to predict n-values at once:

  1. Use model 'predict the next value': make a prediction for the next step, then use it and make a new prediction and so on. An advantage here: 1 model. Disadvantage: your error will grow very fast because your errors are combined on each step.
  2. Use 30 models to predict 30 values (1 step ahead, 2 steps ahead). It is better, but tuning may take a very long time. As suggested here, https://github.com/facebook/prophet may help you in it.
  3. Different combinations.

BTW There is a good answer: https://stats.stackexchange.com/a/354849/183391

But first of all, check your data, interpolate if necessary (both ways), think about nan (if you have). As I see, you have to spend much more time on the data processing, not on model fitting.

And, do not forget, time series are sometimes predictable with linear regression and other models too! According to the shape of the plot it also can be a good approach!

Community
  • 1
  • 1
avchauzov
  • 1,007
  • 1
  • 8
  • 13
0

FBProphet Library from Facebook could be adapted to your needs.

This is a strong library based on Trend + Seasonality + Holidays + Error

This also have an automatic changepoint trend detection and outlier detection.

LaSul
  • 2,231
  • 1
  • 20
  • 36