notes.md

In krzjoa/torchts: Time series Models with torch

TODO

• handling missing data as a separate vignette
• add a vignette with name explanation (index, key, outcome = target etc.)
• monitor hidden state https://fairyonice.github.io/Understand-Keras%27s-RNN-behind-the-scenes-with-a-sin-wave-example.html

• sine https://towardsdatascience.com/can-machine-learn-the-concept-of-sine-4047dced3f11 https://mourafiq.com/2016/05/15/predicting-sequences-using-rnn-in-tensorflow.html https://goelhardik.github.io/2016/05/25/lstm-sine-wave/ https://codesachin.wordpress.com/2016/01/23/predicting-trigonometric-waves-few-steps-ahead-with-lstms-in-tensorflow/

• check model "confusion", i.e. variance for nearly the same examples https://stats.stackexchange.com/questions/220307/rnn-learning-sine-waves-of-different-frequencies

• verbosity level as percentage

Which name?

• h, horizon, prediction_length (modeltime.gluonts)
• timesteps, steps, lookback_length (modeltime.gluonts)
• targets, outcomes <= https://recipes.tidymodels.org/articles/Roles.html
• predictor, inputs
• key (fable), id (modeltime)
• index (fable), date_var (timetk) <= https://tsibble.tidyverts.org/reference/tsibble.html

Conditional RNN

• cond_rnn
• Adding Features To Time Series Model LSTM

Missing values

• See: https://www.tensorflow.org/guide/keras/masking_and_padding
• https://www.nature.com/articles/s41598-018-24271-9

Demand

• https://www.researchgate.net/publication/298725275_A_new_metric_of_absolute_percentage_error_for_intermittent_demand_forecasts
• https://frepple.com/blog/demand-classification/

Metrics

• https://datascience.stackexchange.com/questions/55388/tweedie-loss-for-keras

Libraries

• https://github.com/zalandoresearch/pytorch-ts
• https://pytorch-widedeep.readthedocs.io/en/latest/index.html

Other

• AR-Net

Heuristics

https://stats.stackexchange.com/questions/181/how-to-choose-the-number-of-hidden-layers-and-nodes-in-a-feedforward-neural-netw

R + Keras

https://www.datatechnotes.com/2019/01/regression-example-with-lstm-networks.html

RNN

• A Formal Hierarchy of RNN Architectures https://arxiv.org/abs/2004.08500

Formula proposals

h(value, 7) ~ b(value, 28) + price + lead(price, 5)

Visualization

https://github.com/tensorspace-team/tensorspace

krzjoa/torchts documentation built on June 24, 2022, 5:30 a.m.