Embracing Uncertainty in Forecasting: Enhancing Decision-Making with GluonTS’s Probabilistic Time Series Models | by Everton Gomede, PhD | Apr, 2024

Introduction

In the dynamic field of time series forecasting, the introduction of GluonTS has marked a significant evolution in the tools available to researchers and industry practitioners. As an open-source library built on the robust Apache MXNet framework, GluonTS offers tools that empower users to develop sophisticated, probabilistic forecasting models to handle the uncertainty and variability inherent in time-sensitive data.

Predicting the future isn’t magic, it’s artificial intelligence.

Background

GluonTS is an open-source probabilistic time series modeling library built on the Apache MXNet deep learning framework. It is designed to develop and deploy scalable and accurate time series models. This makes it particularly useful for tasks like forecasting, where understanding uncertainty and variability is crucial.

Critical features of GluonTS include:

1. Probabilistic Forecasts: It specializes in providing point forecasts and distributions that quantify future uncertainty.
2. Built-in Models: The library includes several state-of-the-art models for time series forecasting, such as DeepAR, Transformer-based models, and classical statistical approaches like ARIMA.
3. Customizability and Extendibility: Users can easily define custom models and leverage deep learning components to enhance performance.
4. Datasets and Evaluation Utilities: GluonTS provides utilities for loading and handling time series datasets and tools for rigorously evaluating model performance.

It’s well-suited for academic researchers in machine learning or statistics and industry practitioners working on complex forecasting problems. Whether you’re new to time series analysis or an experienced practitioner, GluonTS provides robust tools to help build effective forecasting models.

Core Strengths of GluonTS

Probabilistic Nature: One of the most critical advancements brought by GluonTS is its focus on probabilistic forecasting. Unlike traditional models that output point estimates, GluonTS models generate probability distributions, which are invaluable for risk management and decision-making processes. This approach provides a fuller picture of potential future outcomes, enabling businesses to prepare for various scenarios and manage risks more effectively.

Rich Model Zoo: GluonTS simplifies the application of complex machine learning models to time series data by providing pre-built implementations of cutting-edge models. This includes the likes of DeepAR, Transformer models, and even classical approaches like Seasonal ARIMA. For practitioners, this means access to state-of-the-art technology without the need to delve into the intricacies of model construction and validation.

Customization and Flexibility: The library’s architecture is comprehensive and highly flexible, accommodating the development of custom models. Practitioners can integrate existing deep learning layers and structures or experiment with novel architectures. This adaptability makes GluonTS particularly appealing in finance and retail, where bespoke solutions are often necessary to address unique forecasting challenges.

User-Friendly Data Handling and Evaluation: GluonTS excels in its user-friendly data handling and model evaluation approach. The library provides utilities for loading and processing time series data, accelerating the model development cycle. Additionally, robust evaluation tools allow for detailed performance analysis, ensuring accurate and reliable models before deployment.

Practical Applications of GluonTS

GluonTS is adept at managing market unpredictabilities in the financial sector through its advanced forecasting models. Financial institutions leverage their probabilistic outputs to quantify potential risks and returns, aiding investment decision-making and risk assessment.

Retail businesses benefit from GluonTS by improving inventory management through accurate demand forecasting. Predicting fluctuations in consumer demand helps optimize stock levels, reducing overstock and stockout situations.

In energy management, GluonTS supports load forecasting, crucial for balancing supply and demand in real-time. Energy companies can forecast consumption patterns more accurately, enhancing operational efficiency and sustainability.

Challenges and Considerations

While GluonTS is a powerful tool, practitioners must know the computational requirements of training complex models, especially on large datasets. Adequate hardware and efficient data handling strategies are crucial for harnessing the full potential of the library.

Moreover, while informative, the probabilistic forecasts provided by GluonTS can be complex to interpret and integrate into decision-making processes, especially for stakeholders unfamiliar with statistical concepts.

Code

Below is a comprehensive example using GluonTS for time series forecasting. This script covers the creation of a synthetic dataset, feature engineering, model selection with hyperparameter tuning, cross-validation, evaluation metrics, and result plotting. The example uses the DeepAR model, a popular choice in GluonTS.

import pandas as pd
import matplotlib.pyplot as plt
from gluonts.dataset.pandas import PandasDataset
from gluonts.dataset.split import split
from gluonts.torch import DeepAREstimator
# Load data from a CSV file into a PandasDataset
df = pd.read_csv(
"https://raw.githubusercontent.com/AileenNielsen/"
"TimeSeriesAnalysisWithPython/master/data/AirPassengers.csv",
index_col=0,
parse_dates=True,
)
dataset = PandasDataset(df, target="#Passengers")
# Split the data for training and testing
training_data, test_gen = split(dataset, offset=-36)
test_data = test_gen.generate_instances(prediction_length=12, windows=3)
# Train the model and make predictions
model = DeepAREstimator(
prediction_length=12, freq="M", trainer_kwargs={"max_epochs": 5}
).train(training_data)
forecasts = list(model.predict(test_data.input))
# Plot predictions
plt.plot(df["1954":], color="black")
for forecast in forecasts:
forecast.plot()
plt.legend(["True values"], loc="upper left", fontsize="xx-large")
plt.show()

Breakdown of the Code:

1. Synthetic Dataset: Create a dataset that simulates a time-varying pattern that is affected by random noise.
2. Feature Engineering: Adds a simple dynamic feature (day of the week).
3. Model Definition: Configures the DeepAR model with custom hyperparameters.
4. Training: Trains the model on the prepared dataset.
5. Cross-Validation: Utilizes the entire dataset for backtesting by simulating a real-world forecast.
6. Evaluation: Uses the built-in GluonTS evaluator to generate metrics that measure forecast accuracy.
7. Plotting: Visualizes forecasts against actual values to provide a clear, visual representation of the model’s performance.

This script assumes the necessary libraries are installed. If not, you’ll need to install them using pip (pip install mxnet gluonts matplotlib pandas numpy).

INFO: 
| Name  | Type        | Params | In sizes                                                         | Out sizes   
------------------------------------------------------------------------------------------------------------------------
0 | model | DeepARModel | 25.9 K | [[1, 1], [1, 1], [1, 1142, 4], [1, 1142], [1, 1142], [1, 30, 4]] | [1, 100, 30]
------------------------------------------------------------------------------------------------------------------------
25.9 K    Trainable params
0         Non-trainable params
25.9 K    Total params
0.104     Total estimated model params size (MB)
INFO:lightning.pytorch.callbacks.model_summary:
| Name  | Type        | Params | In sizes                                                         | Out sizes   
------------------------------------------------------------------------------------------------------------------------
0 | model | DeepARModel | 25.9 K | [[1, 1], [1, 1], [1, 1142, 4], [1, 1142], [1, 1142], [1, 30, 4]] | [1, 100, 30]
------------------------------------------------------------------------------------------------------------------------
25.9 K    Trainable params
0         Non-trainable params
25.9 K    Total params
0.104     Total estimated model params size (MB)

This graph appears to visualize time series data, forecasted values, and confidence intervals. Here’s a breakdown of its components:

1. True Values (Black Line): This represents the observed data points from 1954 to just before 1960.
2. Forecasted Values and Confidence Intervals (Colored Areas): The shaded areas from around 1958 to 1961 represent forecasted values with their associated confidence intervals. Each shaded forecast is for a different time horizon or model, as the other colors indicate.
3. Confidence Intervals (Shaded Areas): The lighter shaded areas around the colored lines indicate the range of values within which the forecast is expected to fall with a certain level of confidence (typically 95%). The wider the shaded area, the greater the uncertainty in the forecast.
4. Trend: The data exhibits a clear upward trend and possibly some form of seasonality, as indicated by the regular patterns in the peaks and troughs.
5. Model Performance: The model captures the overall trend of the data quite well, as the forecasts start within the range of the true values. However, there are deviations, especially noticeable in the green estimates towards the end, where the model forecast deviates significantly from the trend of the true values.
6. Seasonality: The forecast captures the seasonality to some extent, as evidenced by the peaks and troughs of the forecasted intervals following the pattern of the true values.

In summary, the model used to generate these forecasts reasonably captures the overall trend and seasonality of the data, although with some error margins, as expected in any forecast. The graph provides a visual representation of how well the model predictions align with the actual data and can be used to assess the reliability and accuracy of the forecasting model.

Conclusion

GluonTS represents a transformative tool in the arsenal of time series forecasting, offering a blend of advanced probabilistic models, user-friendly features, and customization flexibility. For practitioners across industries, it facilitates informed decision-making under uncertainty, enhances operational efficiencies, and fosters innovation in predictive analytics strategies. As businesses navigate an increasingly data-driven world, tools like GluonTS will be pivotal in leveraging time series data for strategic advantage.

We’ve delved into the intricacies of probabilistic forecasting with GluonTS, exploring its potential to revolutionize how we approach time series data. We’d love to hear from you: How do you manage uncertainty in your forecasting models, and what impact do you think AI-driven tools like GluonTS could have on your predictive analytics? Share your experiences and insights in the comments below — let’s learn from each other’s journeys in harnessing the power of AI for a clearer view of tomorrow.