Skip to content

Latest commit

 

History

History
141 lines (101 loc) · 6.38 KB

TODO.md

File metadata and controls

141 lines (101 loc) · 6.38 KB
Raw

TODO

save / reload the model

self.serialize() => to_dict every where + json

forecast future values => update datasets + forecast

signal transformation almost done

  • cumulative (integrated) tranformation OK

Cutting into Estim/Valid (NEH , NVH) => cross-validation

  1. ratio_E = 0.5
  2. N_train = N_Estim + N_Valid + N_test (=H)
  3. ~~N_Estim = ratio_E * (N_train - H) ~~

~~prediction intervals

  1. ~~improve plotting ... shaded area around prediction intervals neeed prediction intervals first

~~use cross validation

  1. http://robjhyndman.com/hyndsight/tscvexample/
  2. book : https://www.otexts.org/fpp
  3. https://www.otexts.org/fpp/2/5

Residual diagnostics

https://www.otexts.org/fpp/2/6

other trends

  1. Exponential smoothing
  2. exogenous variables OK.
  3. moving average(N). OK
  4. moving median(N). OK

other cycles

1. seasonal
2. user holidays etc (external tables?)

other AR

  1. ARX OK.
  2. VAR ?
  3. order control (look at timedelta ??).

Configuration (Options)

  1. activate/disable transfromations/models/decomposition. OK
  2. configure trends OK.
  3. configure cycles (CycleLength = ?). OK
    • cycle length should be in [5, 7, 12, 24 , 30, 60]
  4. configure ARs (p = ?) OK.
  5. processing : threads etc OK.

Benchmarking

  1. MComp OK.
  2. NN5 OK.
  3. NN3 OK.
  4. Yahoo stocks OK.

speedup things

  1. python is sloooooooooooow (cython ?)
  2. multiprocessing seems OK

timedelta adaptive estimation

  1. allow user control. OK.
  2. truncate timedelta to the nearest unit. OK.
  3. avoid saturday/sunday if not present in the dataset.

LOGGING

~~cross validation for time series

=> http://robjhyndman.com/talks/MelbourneRUG.pdf

~~feature selection (remove unnecessary lags and exogenous variables).

===> smaller model => smaller SQL code !!!

~~real-life examples :

http://stackoverflow.com/questions/10302261/forecasting-time-series-data ~~~~ OK

better graphics

https://stanford.edu/~mwaskom/software/seaborn/. Let someone else do that !!!

~~GitHub Topics

autoregressive benchmark cycles data-frame exogenous forecasting heroku hierarchical horizon jupyter machine-learning-library pandas restful-api scikit-learn seasonal sql sql-generation time-series trends

Forecast Competition

http://eem2017.com/program/forecast-competition

In cooperation with our technical sponsor, we will provide you with a set of different weather input factors, e.g. wind direction, with which you are to forecast the power generation of a wind power plant portfolio. You may participate individually or as a team. The data input is organised in a realistic setting.

2021-08

  1. Prediction Interval Quality
  2. Multiplicative Decompositions (log transform ?). 2022-05-09 OK (#178 : #178)
  3. PyTorch : 2022-05-09 OK (#199 : #199)

2023-07-14

Outliers in Time Series

  1. Outliers detection. https://otexts.com/fpp2/missing-outliers.html
  2. In ARX Models, category 1 : data quality / wrong input. Can be removed. https://otexts.com/fpp2/regression-evaluation.html
  3. In ARX Models, category 2 : Natural, Simply different data. Should not be removed. https://otexts.com/fpp2/regression-evaluation.html
  4. X11 decomposition : The process is entirely automatic and tends to be highly robust to outliers and level shifts in the time series. https://otexts.com/fpp2/x11.html
  5. STL decomposition : outliers may affect the remainder component. https://otexts.com/fpp2/stl.html
  6. Referenced in #230
  7. Outliers removal : estimate the trends/cycles/AR models with an estimation dataset that does not contain the outliers. Can be capped without removal ?
  8. Outliers Reporting. Scatter Plots ? https://otexts.com/fpp2/scatterplots.html
  9. Generic view (not only for time series): https://en.wikipedia.org/wiki/Outlier
  10. Detection method : Tukey's fences, based on measures such as the interquartile range. range = [Q1 - k * (Q3 - Q1) , Q3 + k * (Q3 - Q1)], k > 0. Simple, non-parametric, robust.
  11. Flags/forecast outputs : Tukey uses k = 1.5 to flag as "outlier" and k=3 to flag as "far out".
  12. k value can be used as a training option to remove more-or-less outliers. Default : 3 ?
  13. Interquartile range (IQR = Q3 - Q1): middle 50%. https://en.wikipedia.org/wiki/Interquartile_range
  14. IQR is a Robust measure of scale : for N(0, sigma) , IQR = 1.349 sigma. https://en.wikipedia.org/wiki/Robust_measures_of_scale
  15. Add more candidates with and without oultiers (ARX and ARXO)? Not sure. KISS.
  16. tsoutliers : The tsoutliers() function in the forecast package for R is useful for identifying anomalies in a time series. Excellent !!! http://cran.r-project.org/web/packages/tsoutliers/
  17. tsoutliers : This package implements a procedure based on the approach described in Chen and Liu (1993) for automatic detection of outliers in time series. Innovational outliers, additive outliers, level shifts, temporary changes and seasonal level shifts are considered
  18. Joint Estimation of Model Parameters and Outlier Effects in Time Series. Chung Chen & Lon-Mu Liu. Journal of the American Statistical Association Volume 88, 1993 - Issue 421
  19. http://cran.r-project.org/web/packages/tsoutliers/tsoutliers.pdf
  20. tsoutliers hicp dataset : Harmonised indices of consumer prices in the Euro area.
  21. tsoutliers ipi dataset : Industrial production indices in the manufacturing sector of European Monetary Union countries. 24.tsoutliers bde9915 dataset : seasonal outliers ? Kaiser, R., and Maravall, A. (1999). Seasonal Outliers in Time Series. Banco de España, Servicio de Estudios. Working paper number 9915.
  22. forecast package (tsoutliers function, clean.R): Custom processing for outliers. Iterative MSTL + Smoothing trends + IQR + linear interpolation. https://robjhyndman.com/hyndsight/tsoutliers
  23. fpp2 Gold price dataset : The gold price data contains daily morning gold prices in US dollars from 1 January 1985 to 31 March 1989.

Custom Performance Measures

Add the possibility to customize the performance measures over the given horzion.

One may want to optimize weekly forecasts on a daily model (not take into account small changes within the same week and take forecast aggregates on 7 days).

https://arxiv.org/pdf/2310.10688 A DECODER-ONLY FOUNDATION MODEL FOR TIME-SERIES FORECASTING