164 - An introduction to time series forecasting - Part 4 Using feed forward neural networks

[@victorcallejasfuentes]

All your videos in this serie are so great!! Thank you very much !! Best whishes

[@DigitalSreeni]

Thank you too!

[@vzinko]

Scaling needs to happen after the train test split in order to prevent data leakage from test (i.e. unseen) into train. You fit the scalar on the train and then use that same scalar to transform both train and test sets

[@christophbrand9015]

Thank you :-) You deserve much more views !

[@jurischaber6935]

Again thanks. Great explanation.

[@aination7302]

When you fit the model, please set the Shuffle argument to False.

[@DigitalSreeni]

Thanks for the suggestion.

[@vincentkonate]

I have seen this peculiar result of the prediction being very close to the actual data but shifted of 1 unit in many demonstrations . I think the model simply assume that the next value being equal to the current value is the most accurate prediction.

[@changliu7553]

Great and clear video. Thanks. I do have a question. That is - "what is the most compelling reason for a time series exercise"? what I mean is that a human eyeball can easily see and read trend and period - it takes a lot of time and money to get it "accurate". White noise? no one can do anything about it. So far I have not seen a killer application where ARIMA/CNN/whatever is really crucial. If the analysis can reveal hidden features, great! if it can predict stocks in future, great! the most compelling use is Uber resource management and Amazon logistics planning. Would you agree to some of this?

[@1911maxhtec]

Excellent Video.! Thank you.

[@DigitalSreeni]

You are welcome!

[@ВиталийПлатунов-к6е]

Looks like I finally found a key for my phd) thank you!

[@AbdusSamad-ts7yg]

Very nicely explained Sir

[@fathimahida8878]

Can you do the forecasting for this ?

[@aqilahrose]

seriously sir, this was excellent. Thank you so much.! wondering if you can do a tutorial like this for multivariate time series.

[@DigitalSreeni]

Yes, I will. You should see it in a few weeks, may be by Dec. 8th.

[@khoitruong5923]

Thank you for the best explanation. I have question can we use FFNN for multivariate time series prediction? and how to define the input for it

[@zainabkhan5859]

Thank you for the best explanation👍🏻. Please make videos on automatic sensor dataset with deep neural network. It will help us a lot 🙂

[@jailoslubinda9269]

Thanks for the videos. I just started waxtching the time series parts. A quick question, what models would best be for multiple data points forecasting? I mean, say a dataset for 100 districts of a country with 7 years of monthly data, and missing whole year 6. How can you approach this or can you suggest suitable models for a start? Thanks again

[@junaidlatif2881]

Did you find the answer? Please. Share your answer here!

[@wcottee]

Great video. I can't speak for everyone else but a figure would have been helpful to make the connection between the Neural network structure and the inputs to the KERAS functions. It seems like we have 5 input nodes and a single hidden layer of 64 nodes and a single output node...is this correct? If we would have kept the second dense layer of 32 then we would have 5 input nodes and two hidden layers, the first of 64 nodes and the second of 32 nodes (then followed by the single output node of course).

[@ipekuyguner5102]

do you have recommondation for non stationary data normalisation? Bcause I suppose here we do minmax according the whole data (including test data ). But we should not use anything from test statistics ? for stationary datasets, maybe it does not cause big problem but for unstationary datasets, isn't it data leaking?

[@DigitalSreeni]

Try adaptive normalization for non-stationary data. www.researchgate.net/profile/Gisele_Pappa/publication/221532708_Adaptive_Normalization_A_novel_data_normalization_approach_for_non-stationary_time_series/links/0c96051cc596087da6000000/Adaptive-Normalization-A-novel-data-normalization-approach-for-non-stationary-time-series.pdf

[@manisathya9108]

Appreciate for your great video .. could you please advice how to add future prediction like 30 points in future - is it possible

[@himanshutyagi4970]

Which model can I use if I have multiple small multivariate time series data from a system ? Thanks

[@yogeshdeotale7]

hello sir .... why you haven't use standard scaler instead of min max scaler?

[@WahranRai]

Your testing / validation are not O.K. If you you want to simulate time series you must respect the sequence. Perform a back testing loop : test the first, adding it to training set, fitting and set the next etc.... The training test is expanded dynamically by the testing set one instance by one instance

[@DigitalSreeni]

Thanks for the comment Wahran. I will have to go back and verify what I have done wrong and correct it in my code. Thanks.

[@LesleyPotters]

Thanks for the good video. Could you please elaborate on the choice of the sequence size? Why is this 5 months? Would it be possible to (additionally) include 12 months time steps? Thanks!

[@shuvoislam83]

What this RMSE value is telling about our model? What is the range of the RMSE for this model? Could you please explain a bit?

[@laquidakristyne6952]

Thank you for your video! Can I put it this way that you transformed a time series forcasting problem into a regression problem? And another question is, how to change the model to perform multi-step-ahead predictions?

[@jaikishank]

Thanks for an excellent video. I have one small clarification. If we have the statement under def to_sequences(dataset, seq_size=1): for i in range(len(dataset)-seq_size-1): this does not take the last y value. The operations stop at the penultimate row. Kindly correct me if my understanding is wrong.Instead, we can have for i in range(len(dataset)-seq_size): which considers the last record.

[@kalymdros]

For example, if the window is 5 steps, the index i, starts from 0. When the window reaches the end, lets say 100, then the window will fill the 100, 99, 98, 97, 96 positions (5 steps). This means the index i will be stopped at the 96th position. So the end is 100 - 5 - 1 = 96. So its len(dataset) - seq_size -1. I know this is 2 years late, but I was also thinking about it and there you go.

[@projectsbysrikanth]

Do we need to remove seasonality from this kind of data? Will it help and have you covered it in any other lesson? (Great video series btw)

[@IMRANKHAN-lc7vx]

First of all thank you for your excellent videos. I had one question: Which ML technique to use if we have input(Set of parameters and Images) and output (A time step series of 3D images - Very large file). Kindly can you please give me a suggestion?

[@hidayatzulqarnain]

thanks for your video, i learn it for my thesis. but, I got an error in lines 57 - 73 which is " def to_sequence....." it throws "self._engine.get_loc(casted_key) InvalidIndexError: (slice(0, 10, None), 0)" i hope you read my comment and help me to solve the error

[@kirangeorge7659]

Thank you very much for the video, helped me learn a lot. One question I had is: How is 2.3% accuracy on train data a good score?

[@JJGhostHunters]

I followed everything except the "shifting" part at the end.