truesight 0.0.3a1

Truesight is a python package for time series prediction using deep learning and statistical models.

TrueSight ✨

The TrueSight model is a hybrid forecasting tool that uses statistical forecasting models together with a Deep Neural Network (DNN) to make predictions. The TrueSight Preprocessor class is responsible for getting all the statistical forecasters in one place. It can handle forecasters from packages like statsforecast, scikit-learn, pmdarima, and others. You just need to the ModelWrapper Class to standardize the method calls.

All you need to do before using this package, is create a pandas dataframe with the following structure:

• unique_id: A string that uniquely identifies each time series in the dataframe
• ds: A datetime column with the date of each time step. The dates must be in the correct frequency for the date_freq parameter
• y: The values of the time series

and run the steps in the Usage section. Easy peasy! 😎

Instalation 💻

To install the TrueSight package, just run:

pip install truesight

We also recommend installing the statsforecast package for the statistical forecasters:

pip install statsforecast

Usage 🚀

Import the necessary modules

import tensorflow as tf
from truesight.preprocessing import Preprocessor
from truesight.core import TrueSight
from truesight.metrics import Evaluator, smape, mape, mse, rmse, mae
from truesight.utils import AutoTune, generate_syntetic_data

Load the data

num_time_steps = 60
season_length = 12
forecast_horizon = 12
df = generate_syntetic_data(num_time_steps, season_length, 100)

Create and run the preprocessor class. You can include as many statistical models as you need in the models parameter, just make sure to use the ModelWrapper class. However, keep in mind that more models mean longer processing time. It's important to set a fallback model in case any of the informed models fail to fit.

from statsforecast.models import SeasonalNaive, AutoETS
from sklearn.linear_model import LinearRegression
from truesight.models import AdditiveDecomposition
from truesight.utils import ModelWrapper

models = [
    ModelWrapper(LinearRegression, horizon=forecast_horizon, season_length=season_length, alias="LinearRegression"), 
    ModelWrapper(SeasonalNaive, horizon=forecast_horizon, season_length=season_length), 
    ModelWrapper(AutoETS, horizon=forecast_horizon, season_length=season_length),
    ModelWrapper(AdditiveDecomposition, horizon=forecast_horizon, season_length=season_length)
]

preprocessor = Preprocessor(df)
X_train, Y_train, ids_train, X_val, Y_val, ids_val, models = preprocessor.make_dataset(
    forecast_horizon = 12, 
    season_length = 12,
    date_freq = "MS", 
    models = models, 
    fallback_model = ModelWrapper(SeasonalNaive, horizon=forecast_horizon, season_length=season_length),
    verbose = True
    )

Create the model and automatical automatically find the hyperparameters

optimizer = tf.keras.optimizers.Adam
hparams, optimizer = AutoTune(optimizer=optimizer).tune(X_train, Y_train, n_trials = 20, epochs = 10, batch_size = 32, stats_models = models)
ts = TrueSight(models, forecast_horizon)
ts.set_hparams(hparams)
ts.compile(optimizer=optimizer, loss='mse')

Or set then manually

optimizer = tf.keras.optimizers.Adam
ts = TrueSight(models, forecast_horizon, filter_size = 128, context_size = 512, hidden_size = 1024, dropout_rate = 0.1)
ts.compile(optimizer=optimizer, loss='mse')

Train the model, as the model is built on the tensorflow framework, any tensorflow callback can be used

callbacks = [
    tf.keras.callbacks.EarlyStopping(patience = 100, restore_best_weights = True, monitor = "val_loss"),
    tf.keras.callbacks.ReduceLROnPlateau(monitor = "val_loss", factor = 0.5, patience = 25, verbose = False),
]
ts.fit(
    x = X_train, y = Y_train, 
    validation_data = [X_val, Y_val], 
    batch_size = 128, 
    epochs = 1000, 
    verbose = False, 
    callbacks = callbacks,
)
ts.plot_training_history()

Evaluate the results

yhat = ts.predict(X_val, n_repeats = 100, n_quantiles = 15, verbose = False)
evaluator = Evaluator(X_val, Y_val, yhat, ids_val)
evaluator.evaluate_prediction(evaluators = [smape, mape, mse, rmse, mae], return_mean = True)

metric	value
smape	0.234369
mape	0.293816
mse	816.238082
rmse	21.218396
mae	15.885432

evaluator.plot_exemple()