I have an ensemble of 231 time series, the largest among them being 14 lines long. The task at hand is to try to predict these time-series. But I'm finding this difficult due to the very small size of the data. Any suggestions about what algorithm to use? I'm thinking about going for a hidden markov model, but I don't know if that's a wise choice.
-
$\begingroup$ Why do you want to train a model in first place? Will you have more data to retrain your model later? $\endgroup$– pedrumAug 3, 2020 at 19:56
-
1$\begingroup$ @pedrum yes i will. $\endgroup$– Ons BouaradaAug 4, 2020 at 13:23
-
$\begingroup$ Nice. Then don’t treat your data as very small data. Replicate some data by adding very small noise into your data and treat it normally, when you get more actual data, use the same algorithm for it. Like this not only you will have tried to generalise your initial model, but also will have examined your appropriate algorithm with your toy data. $\endgroup$– pedrumAug 4, 2020 at 14:22
-
$\begingroup$ Not sure what ensemble means here. Do these 231 times series represent 231 different processes? For example do you have one time period sampling of the closing price from each of 231 different stocks or do you have 231 time period samplings of one stock? If it is the former you generally use one sampling to predict at a time. $\endgroup$– Brian O'DonnellAug 4, 2020 at 14:47
-
$\begingroup$ What have you tried? $\endgroup$– Brian O'DonnellAug 4, 2020 at 14:49
1 Answer
Of course depends on your type of data, but Holt-Winter models can have different degree of complexity and use moving average, trend, and seasonality. This is most useful if the data is not hierarchical, meaning that the time-series are independent from each other. If time-series are relatives of each other then you can also try aggregating them, predict at aggregate level and then disaggregate. The following can be a good resource: