Consider a lot number of objects each with a time series history. these objects could be clustered to groups based on their history information similariry.
In each group of similar objects, the new arrived changes are dependent. for example in holidays, the values will be decreased and in working days will be increased for all similar objects in a group.
i want to use this data to adapt the prediction of each object in real time. I mean using a normal time series prediction for each object separately in conjunction with an average of values coming in real time from similar objects.
could you help me how to do this?
What is the frequency of series?
Generally, this sounds like a difficult task.
One simple approach to start with is to calculate seasonal indices for each series and use averaged indices to re-calculate forecasts.
To compare alternative approaches you can use this framework for forecasts evaluation and comparison:
Conference Paper Data Formats and Visual Tools for Forecast Evaluation in Cyb...
For more details on error metrics let me recommend this chapter:
Chapter Forecast Error Measures: Critical Review and Practical Recommendations
Hello Babak,
i don't know if I understood your question correctly. Is it about clustering similar series into groups of series? If yes, you should take a look onto the work of Rob Hyndman and the feats package in R which lets you extract up to 48 features of a series. These can then be reduced to principal components or cluster analysis (or a cluster analysis of the PCs).
Best wishes,
--Holger
The first step is visual analysis.
You can use spaghetti plots, e.g., as shown here:
Poster A Visual Framework for Longitudinal and Panel Studies (with ...
This is needed to detect data flaws, identify trends/seasonality, etc.
Then, if you have hourly observations, perhaps you may use trigonometric approximation to describe weekly cycle.
But, of course, if you have national holidays, this will complicate the task. The task will be even more difficult if you have calendar events occurring at different dates (e.g., Easter or Chinese New Year). This will be modeled using dummy variables with some lags included.
Then if you find series where you believe have similar seasonal and holiday effects, you can use the same dummies.
It's also important to consider data transformations if you have non-negative data. E.g., you can apply log-transformation. You can find more on data transformations in this guide:
https://xdplot.blogspot.com/p/data-transformations-guidelines.html
If you use the log-transformation, you can then model additive seasonality and additive dummy variables for calendar events.
In order to evaluate forecasting performance I would recommend using the AvgRelMAE metric because if you back-transform forecasts, they will be optimal under linear loss, and the AvgRelMAE metric is suitable for this case. You can find more on the AvgRelMAE in these chapters:
Chapter Forecast Error Measures: Critical Review and Practical Recommendations
Chapter Measuring Forecasting Accuracy: Problems and Recommendations...
Hi Babak Yazdanpanah ,
Based on whatever little I could understand from your explanation,
1] You have a set of time series data, all having identical or different time axis (at one point you have mentioned hourly data)
2] You want to predict one parameter based on the present and past values of dependent parameters (the availability of these set of dependent parameters is itself a variable).
I have two publications in this respect.
Check if it helps you in any manner.
https://www.researchgate.net/publication/341141896_Week_Ahead_Time_Series_Prediction_of_Sea_Surface_Temperature_Using_Nonlinear_Autoregressive_Network_with_and_Without_Exogenous_Inputs
Conference Paper Dependency Investigation of Sea Surface Temperature on Sea B...
If the data are relevant to what you are doing it will be ok. However, historical data may not be the representative of what is currently happening, in this case more current data are needed.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series