basic Nx1 d array for forecasting

[andrewcztrack]

Hi @petroniocandido I am really interested in the work!
I was wondering if i have a basic nX1 d array, I am trying to work out how to call the train and test function on the data.
would you have small example that you can share.
Best,
Andrew

[petroniocandido]

Hi @andrewcztrack

Thanks for your interest in our work!

I don't know if I correctly understood your question. You have a univariate time series (a np.ndarray with (n,1) shape ? ), and want to know how to split the train and test sets? It is correct in my understanding?

Sincerely yours

[andrewcztrack]

hey @petroniocandido !! yes yes !! thank you!
That is exactly my question. if you could give me a brief example that would be amazing.
:)
Many thanks,
Andrew

[petroniocandido]

Hi @andrewcztrack !

There is a good practice that advises the employment of Sliding Window Cross-Validation. What does this mean?

The sliding window cross-validation is a kind of k-fold cross-validation where the folds (or data windows) are overlapped. The data window contains the train and test splits (whose ratios use to be 70% / 30% of the data). For instance, the initial data window covers the interval at t=[0..100], the second data window will be t=[50,150], and so on. Using this method you will train several models, with different data windows, and measure the average performance (RMSE, MAPE, U, etc) of the model. The sliding window cross-validation is widely used for hyperparameter optimization and fine-tuning of hyperparameters.

Indeed, pyFTS library has a method for sliding window cross-validation:

from pyFTS.common import Util

for window_count, train, test in Util.sliding_window(data, windowsize, train_ratio, inc=increment_size):
     train model
     test model

If you are just interested in fit a simple model, without performing any kind of optimization, just pick your time series e split it into the train and test subsets with 70% / 30% ratios [or 80% / 20%].

The NSFTS model was projected for nonstationary data, where it is known that the train and test splits have different universes of discourses, but not completely different behavior. The rules are learned in the train data represents the behavior of the time series and the main assumption is that the behavior keeps, even outside the original universe of discourse. If the time series is likely to change everything [universe of discourse, behavior, etc] it means a concept drift, and maybe the Retrainer model will be better.

I hope have answered your questions, but if I don't, do not hesitate to keep in touch!

Best regards

[andrewcztrack]

Hi @petroniocandido !! i have a basic example.
does the below example make sense.

from pyFTS.benchmarks import naive
from pyFTS.models.incremental import TimeVariant, IncrementalEnsemble
from pyFTS.models.nonstationary import common, perturbation, partitioners as nspart
from pyFTS.models.nonstationary import nsfts, util as nsUtil
from pyFTS.partitioners import Grid
from pyFTS.models import hofts, pwfts
from pyFTS.benchmarks import Measures
from pyFTS.common import Transformations, Util as cUtil

from pyFTS.benchmarks import Util as bUtil

import yfinance as yf
data = yf.download("SPY", start="2014-01-01", end="2020-12-30")['Adj Close']
dd=data.to_numpy()

window = 100
batch = 10
num_models = 2
l = len(dd)
train = dd[:l//10]
test = dd[l//10:]
nsfs = nspart.simplenonstationary_gridpartitioner_builder(data=train, npart=35, transformation=None)
model = nsfts.WeightedNonStationaryFTS(partitioner=nsfs)

model.fit(train)
#forecasts = model.predict(test)
mse,mape,u = Measures.get_point_statistics(test, model)
row = [key, model.shortname, rmse, mape, u]
rows.append(row)

m = {'rmse': rmse, 'mape': mape, 'u': u}


resultados = pd.DataFrame(rows,columns=["Dataset","Model","RMSE", "MAPE", "U"])
plt.plot(forecasts)
plt.plot(test)
resultados

[petroniocandido]

What do you think to create a google colab notebook with this code and share the link here? I can add more code examples more easily...

[andrewcztrack]

Hi @petroniocandido ! that would be amazing!

https://colab.research.google.com/drive/1ik0WKJsS_ZwWdlGUZ9YYHrfl4oIttwYc#scrollTo=_rjGhlNsEvUs

Can you see if the above works :).
cheers,
Andrew

[andrewcztrack]

Hi @petroniocandido i have added you to google colab notebook :).

Thank you so much for your help!
Best,
Andrew