How would you judge the performance of an LSTM for time series predictions? For an LSTM model for forecasting time series, what metrics or tests would you use to evaluate its performance (i.e loss, RMSE, accuracy etc). I'm slightly confused because I read that time series forecasting is considered a regression problem so accuracy doesn't apply but I have also seen many time series models use accuracy as a metric. Any explanation would be greatly appreciated!
 A: Time series prediction is a regression problem and you should consult to regression error metrics. However, you can make a classification based on time series. For example, you might need to decide a person's health status considering some physiological recordings. Then, your problem turns into a classification task where you can use classification metrics such as accuracy. I presume this is the source of your confusion
There are multiple metrics for quantifying time series prediction models. Some of them:


*

*Mean Absolute Error:  $\frac{1}{N}\sum_{i=1}^{N}|y_i-y_i'|$ 

*Mean Squared Error:  $\frac{1}{N}\sum_{i=1}^{N}(y_i-y_i')^2$ 

*Mean Percentage Absolute Error : $(\frac{1}{N}\sum_{i=1}^{N}|\frac{y_i-y_i'}{y_i}|)$x$100$
where $y,y'$ and $N$ are observed time samples, predicted time samples and total number of time samples used for error calculation, respectively. The procedure is straightforward: using previous time samples, you predict following $N$ time samples and calculate error based on these predicted samples.
Although these all are widely used, percentage error is a good starting point since it is not affected by the data scaling and therefore, easier to interpret.
