I use DeepAR for time series prediction using multiple input timeseries.
Shortly, I have several timeseries (lets say Input_1, Input_2, Input_3 and Input_5 and TARGET). All these values are from same system and have a relation somehow. In addition, these time-series has no periodic behavior, in other words, there is no seasonality for any of the signals. I want to create a model and predictor for TARGET signal. Figure given below helps to explain behavior. (None of them is periodic)
BLUE signals are input for predictor to forecast RED signal. In other words, I want to generate a model to find how RED behaves when BLUE signals are input.
I choose Sagemaker DeepAR for this problem but there is an issue about DeepAR usage. DeepAR requires context_length parameter to define how far in the past the network can see. Why does DeepAR need this parameter for nonperiodic timeseries? There is no periodic behavior and trend as it can be seen on my plot?
Hi All,
I use DeepAR for time series prediction using multiple input timeseries. Shortly, I have several timeseries (lets say Input_1, Input_2, Input_3 and Input_5 and TARGET). All these values are from same system and have a relation somehow. In addition, these time-series has no periodic behavior, in other words, there is no seasonality for any of the signals. I want to create a model and predictor for TARGET signal. Figure given below helps to explain behavior. (None of them is periodic)
BLUE signals are input for predictor to forecast RED signal. In other words, I want to generate a model to find how RED behaves when BLUE signals are input.
I choose Sagemaker DeepAR for this problem but there is an issue about DeepAR usage. DeepAR requires context_length parameter to define how far in the past the network can see. Why does DeepAR need this parameter for nonperiodic timeseries? There is no periodic behavior and trend as it can be seen on my plot?