I would try to include additional indicator features for weekdays or weekends directly into your Arima model.
Also I believe that RNNs are useful mainly for highly non-linear problems. Non-linearities in a problem such as sales forecasting are best handled through interaction terms or by including non-linear transformations (e.g. logarithm) of existing features into your model.
The problem with ARIMA is that it's more of an exponential regression rather than deep learning. We already have added weekday,weekend,month variable to it but its more of regression than deep learning.
So you are saying that RNN won't be suited for non-linear sales forecasting?
I don't have the details of your exact use case, but I cannot imagine any complex non-linearities involved in your sales process. If ARIMA models produce decent results for your use case I would try to improve the ARIMA model through additional data, rather than switching to deep learning.
If you are convinced that there are complex non-linearities that an ARIMA model cannot describe, then I would try to use RNNs to find a pattern in your ARIMA models' residuals and try to augment your ARIMA model with an RNN.
>"I cannot imagine any complex non-linearities involved in your sales process"
Knowing nothing about his sales process and features, I would assume there are many complex non-linearities (to be possibly leveraged for better predictions). I find this statement bizarre.
What I mean by this statement is that there are lots of "tricks" such as interaction terms, regime switching and non-linear transformation of features to handle non-linearities in linear models (e.g. different food sold before Christmas).
But if you can give me an example of a non-linearity in sales forecasting that cannot be fit by a linear model but can by an RNN I'd honestly be really interested in that.
> If ARIMA models produce decent results for your use case I would try to improve the ARIMA model through additional data, rather than switching to deep learning.
ARIMA models cannot take into consideration special scenarios or even multivariate features (Like location and time). It's good for simple forecasting but when you need more "human like" predictions we are betting on neural networks.
For example we would instinctively know that sales of turkey goes up on Thanksgiving. ARIMA model cannot factor this but RNN/LSTM should be able to (Theoretically speaking).
I'd suggest lecture 14 of the fast.ai mooc on some advice on feature engineering for timeseries data to make it possible to model with something like regression (or a non-recurrent neural net - can work surprisingly well)
I read the article and seems to be well-written though lacking.
For even more customized RNNs such as attention mechanism, beam search as in Seq2Seq, you'll need to skip the tf.dynamic_rnn abstraction and use a symbolic loop directly: tf.while_loop
I think that's covered in the article - there's a passage on using `tf.scan` when the `tf.dynamic_rnn` abstraction won't cut it. `tf.scan` is more flexible than `tf.dynamic_rnn`, but provides a little more scaffolding for RNNs than using `tf.while_loop` directly.
scan implements strict semantics so it will always execute the same number of timesteps no matter what the accumulator is (nan).
while_loop implements dynamic execution (quit once cond is not met) and at the same time allows parallel execution when some ops are not dependent on accumulator.
If you read the code for `dynamic_rnn` and contrib.legacy Seq2seq model you'll find while_loop. I have yet to see tensorflow library code using tf.scan anywhere!
Also, internally, scan is defined using while_loop. In my code, I find scan lacking in RNN and always have to fall back to while_loop.
Here is video of a talk by the RNN/Seq2Seq author himself:
I don't follow. tf.scan will execute as many time steps as there are elements in the input series, which is the same behavior you'd get with tf.while_loop or tf.dynamic_rnn. It does not execute for a fixed number of time steps, which I think is what you're implying?
The difference from using tf.while_loop directly is that tf.scan handles the logistics of an accumulator to keep track of hidden states, so you don't have to implement that piece yourself.
As you say, tf.scan uses tf.while_loop internally; it's not particularly different from something you might build using tf.while_loop yourself.
I agree that you have to use a tf.while_loop in those cases. But then tf.scan isn't an option, so I don't understand what you mean by 'quit early' or 'saves time'.
When tf.scan is possible, i.e. when you have an input sequence you want to scan over, it is a perfectly good option.
Do you know if using tf.while_loop speed things up? Using dynamic_rnn at the moment and it's _so_ slow. I'm not finding implementations using tf.while_loop, there's dynamic_rnn as you said but that's so convoluted to read (like TF code..).
We are looking to replace our Arima models with RNNs and the results so far has been far from satisfactory.
The usecase is: based on sale quantity in past year, predict the sale quantity tomorrow.
Regression does not consider weekdays or weekends or similar bumps and we thought RNN w/LSTM would be well suited for this problem