A uni-variate forecasting framework using N-BEATS and RevIN implemented in PyTorch
N-BEATS paper - https://arxiv.org/pdf/1905.10437.pdf
RevIN paper - https://openreview.net/pdf?id=cGDAkQo1C0p
N-BEATS model is one of the latest neural networks developed with a focus on univariate times series point forecasting problem.
RevIN is a normalization method which addresses the distribution shift problem in time series data when working on such data through neural networks.
RevIN normalizing-denormalizing function is implemented using official PyTorch implementation of RevIN (https://github.com/ts-kim/RevIN).
Which takes Torch.tensor input and return normalized/denormalized Torch.tensor output
N-BEATS model is implemented using darts library (https://github.com/unit8co/darts) which is sourced from PyTorch Lightning implementation (https://github.com/Lightning-AI/lightning)
assignment.ipynb has all other important info in comments if needed
BENCHMARKS
| TIMESERIES | MAE | MSE | RMSE | MAPE |
|---|---|---|---|---|
| 0 | 0.12612509900170069 | 0.024198295009439465 | 0.1555580117172994 | 0.26290317501470517 |
| 1 | 0.0655551819072414 | 0.010388990936702822 | 0.10192639960629837 | 0.22990608456039982 |
| 2 | 1.1617019888826996 | 2.037466163857422 | 1.4273983900290144 | 2.655545056665787 |
| 3 | 70.48688230897851 | 6553.408803521374 | 80.9531272003829 | 2.032005256858877 |
| 4 | 1.4084744606521534e-13 | 3.1115688126098876e-26 | 1.7639639487840696e-13 | |
| 5 | 3.039139617782327 | 20.88048603587931 | 4.569517046240151 | 72.60391462214513 |
| 6 | 8488952128.244324 | 1.363662944499226e+21 | 36927807198.63049 | 344756319.4946807 |
| 7 | 140.6379736135317 | 35770.9602962864 | 189.13212391417383 | 10.179926714459999 |
| 8 | 1877.5885559776436 | 7756569.713358287 | 2785.061886809391 | 30.018728996772957 |
| 9 | 3542.308605455195 | 24286868.004587237 | 4928.170857893143 | 47.777416674041355 |
| 10 | 0.07024170650305607 | 0.014467931671487454 | 0.12028271559741015 | 1.0320939669181088 |
| 11 | 0.047742708190771414 | 0.00332314210104095 | 0.057646700695191136 | 0.5294343627484632 |
| 12 | 5852.87252414172 | 58722886.830946505 | 7663.085986138124 | 22.038329867060696 |
| 13 | 2886068.541248724 | 18774410752400.434 | 4332944.812988095 | 38312716.134632975 |
| 14 | 0.12600381018542214 | 0.05761956544201853 | 0.24004075787669588 | 0.12662619979557901 |
| 15 | 0.48914169354784387 | 0.4376590475506016 | 0.6615580454885283 | 0.49508263042771716 |
| 16 | 0.35548380074414565 | 0.20983156205781772 | 0.4580737517669155 | 0.39161314440144396 |
| 17 | 0.07098968135003413 | 0.009550072694588385 | 0.09772447336562314 | 1.4631623618955127 |
| 18 | 6853.632921015647 | 85478101.2273193 | 9245.436778612426 | 8.159490551497283 |
| 19 | 0.05952956009538465 | 0.005224873333539411 | 0.07228328529846587 | 0.35106847251103346 |
| 20 | 9.145265839370174 | 461.2808902593384 | 21.477450739306523 | 7.73734713066623 |
| 21 | 33888.15927370629 | 1350720499.1972575 | 36752.149586075335 | 0.31759162330177126 |
| 22 | 4064.2374801167884 | 32610093.981285725 | 5710.524842891915 | 30.340762242132595 |
| 23 | 3514.6451818059686 | 25725087.22986155 | 5071.990460348043 | 10.958529558697904 |
| 24 | 0.5092879996255936 | 0.5891496655281042 | 0.7675608546090038 | 0.5186679522491247 |
| 25 | 35.65531101621262 | 1981.1184289923615 | 44.50975655957199 | 81.83476949799166 |
| 26 | 1118.0680364597617 | 2765489.174728455 | 1662.9759994445064 | 9.372476060502413 |
| 27 | 3.039139617782327 | 20.88048603587931 | 4.569517046240151 | 72.60391462214513 |
| 28 | 115146.50216348954 | 17279073382.760998 | 131449.88924590618 | 3.7293221804651027 |
| 29 | 106523.21814985787 | 14929595570.926537 | 122186.72420081707 | 3.2847897210555836 |
| 30 | 1.2878560378821422 | 2.561985239564101 | 1.6006202671352443 | 14.891567160549412 |
| 31 | 2928.852251371396 | 24809561.741004337 | 4980.91976857732 | 17.277042413999983 |
| 32 | 1877.5885559776436 | 7756569.713358287 | 2785.061886809391 | 30.018728996772957 |
| 33 | 0.6087130993970967 | 1.7965158762459685 | 1.340341701300817 | 22.73870113793243 |
| 34 | 20404.106006947026 | 853029046.2680435 | 29206.660991425288 | 2.3581181853944764 |
| 35 | 42.20025870844651 | 3408.786101619876 | 58.38481053852856 | 3.8903347774709482 |
| 36 | 0.08552587518235034 | 0.009584039547589447 | 0.09789810798779232 | 0.5779657900812831 |
| 37 | 0.7638648917715476 | 0.6980849872722453 | 0.8355148037421272 | 2.2877074484301305 |
| 38 | 0.3739558429113322 | 3.6857340596412254 | 1.9198265701987838 | 0.4242459536316819 |
| 39 | 0.2812693863413277 | 0.18947158261781974 | 0.43528333602128594 | 0.283267534754295 |
| 40 | 3732.9162466036564 | 25376494.43959771 | 5037.508753302342 | 24.99884629815229 |
| 41 | 15.790168293586634 | 453.8594364422328 | 21.30397700999118 | 12.899610341308426 |
| 42 | 1.1391007073283126 | 1.838257519564165 | 1.3558235576815167 | 11.66368116697407 |
| 43 | 0.2517552394899248 | 0.10484285402094216 | 0.3237944626162439 | 14.53742038865659 |
| 44 | 0.32810279096666933 | 0.14863118128095604 | 0.38552714726845894 | 1.125913704123885 |
| 45 | 9540720819.294909 | 1.3523454091709852e+21 | 36774249267.26561 | 35785906.6392745 |
| 46 | 424.1707646458827 | 352128.21391247964 | 593.403921382796 | 56.435301858395306 |
| 47 | 0.37601644374966614 | 0.18629205802382057 | 0.4316156369083731 | 2.343597518548591 |
| 48 | 1253.6903027227097 | 3045104.6537922765 | 1745.0228232869267 | 8.097194394930485 |
| 49 | 2595924.0578850713 | 15922715349168.938 | 3990327.7245320263 | 5282933.921404671 |
ISSUES [AND SOLUTIONS]
hyperparameters are tuned manually [NEED TO IMPLEMENT CUSTOM CLASS FOR HYPER-PARAMETER TUNING]
interpretable NBeats model which may detect seasonality and trend when there is none (0 or constant values) [USING ENSEMBLE TECHNIQUES WITH GENERIC MODELS MAY SOLVE THE PROBLEM]
HOW TO USE REPO
download zip and extract assignment.ipynb file and dataset50 folder
create a folder containing dataset folder (named dataset50) and an empty forecast images folder (named forecasts)
give this path to ASSIGNMENT_PATH variable in the ipynb file
the code will automatically read from the dataset folder and store forecast plots to the forecasts folder
at the end of the complete execution, it will also store all the error metrics of the timeseries' in the dataset, in a csv file named metrics.csv which can be found in the main folder