Seasonality and Demand Forecasting with Keras
Vložit
- čas přidán 27. 12. 2022
- In this video we see how to remove trend and seasonality from a dataset prior to predicting with Keras. Neural networks do not handle trend or seasonality very well without pre/post-processing.
Kaggle Dataset:
www.kaggle.com/datasets/jeffh...
Code for this video:
www.kaggle.com/code/jeffheato...
Kaggle Community Competition for this Dataset:
www.kaggle.com/competitions/a...
Demand Forecasting Playlist
• Demand Forecasting wit...
~~~~~~~~~~~~~~~ CONNECT ~~~~~~~~~~~~~~~
🖥️ Website: www.heatonresearch.com/
🐦 Twitter - / jeffheaton
😸🐙 GitHub - github.com/jeffheaton
📸 Instagram - / jeffheatondotcom
🦾 Discord: / discord
▶️ Subscribe: czcams.com/users/heatonresea...
~~~~~~~~~~~~~~ SUPPORT ME 🙏~~~~~~~~~~~~~~
🅿 Patreon - / jeffheaton
🙏 Other Ways to Support (some free) - www.heatonresearch.com/suppor...
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
#Python #Tensorflow #Keras #csv #png #jpg #csv - Věda a technologie
3 minutes in and already worth the watch. Thanks Jeff!
I've learned that feeding a sufficiently complex model the day_of_year, training it on 2 years and predicting for the last year, you will end up with basically the mean value of the series per day_of_year. It won't have much predictive power. Also it is very important to predict for an entire period of the seasonality with the lowest frequency, otherwise your detrending operation could be distorting your evaluation metrics.
The real goal is to learn the influence of external_regressors and differentiate them from the autoregressive components.
I would really love it if you compared the complex method to a very simple method (like moving window groupby mean on datetime features and use as prediction, or even just shifting the data the required amount to get a prediction). It would also be interesting if you could discuss metrics we could use to determine if it is worth the trouble of training a very complex model that requires a complex pipeline and lots of data to continue functioning.
I've been watching for years and have you to thank much of my academic success, so in no means do i mean the above as criticism, more thinking out loud :).
Much respect to your work, Jeff 👍🏻
This is such valuable information, thank you so much :D always look forward to new videos from you!
Oh thanks, so glad it was useful.
very useful, thank you Jeff.
Great to see this.
Good afternoon, Jeff. I am Dmitry from St. Petersburg. I use your encog 3.4 library in java in time series forecasting. Your wonderful work! I wanted to ask you a few questions if you have time. I use RPROP for training.
1 Why is it that when there is one internal layer in the network, the training goes through many iterations. But when there is more than one, then very soon the error stops decreasing?
2 Which step of delta error reduction should be considered as a signal for the end of training?
3 I'm using java 11, can the library work incorrectly because of this?
Thanks.
I am stuck in 18th step, how can I get help?
ok, so why is it that we can train a RNN to learn a sine wave but not to remove seasonality?
first conv layer is 1D, you said 2D @7 min but you explained 1D :)