Forecasting with Elixir: A Look at Sser and Neuro-Prophet for Time Series Analysis

• 0:00:25 Introduction: The speaker begins by discussing the challenges of forecasting and introduces the topic of using Elixir for time series analysis.
• 0:01:09 Zillow's Losses: The example of Zillow losing over $300 million due to a faulty data science model is presented, highlighting the importance of accurate forecasting.
• 0:01:45 Introducing the Speaker: The speaker introduces himself, his background, and his current role at CloudWalk, a global fintech company.
• 0:03:45 Neuro-Prophet and Prophet: The speaker discusses Neuro-Prophet, a neural network-based forecasting tool, and its predecessor, Prophet, developed by Facebook.
• 0:06:30 Understandability and Explainability: A key focus of both Neuro-Prophet and Prophet is the ability to understand and explain the models to stakeholders.
• 0:07:24 Decomposing Time Series: Neuro-Prophet decomposes time series data into trend, seasonal effects, holidays, and regressions for better analysis.
• 0:09:07 Trend Component: The trend component is explained as a linear trend between segments, with Neuro-Prophet using the last segment for future forecasting.
• 0:09:58 Seasonality as a Continuous Function: Seasonality is treated as a continuous and periodic function, decomposed into Fourier series (signs and cosines).
• 0:11:04 Autoregressors and Lag Regression: The concepts of autoregressors (using past values to predict future values) and lag regression (using the effect of known past variables) are introduced.
• 0:13:33 Holiday Effects: The importance of considering holiday effects in business settings is highlighted, with Neuro-Prophet offering a way to decompose these effects.
• 0:14:09 SourceLevel and Engineering Metrics: The speaker shares his experience using Neuro-Prophet at SourceLevel to predict engineering metrics and identify bottlenecks.
• 0:15:14 Introducing Sser: The speaker introduces Sser, an Elixir library for time series forecasting inspired by Neuro-Prophet.
• 0:16:39 Sser's Current Capabilities: Sser can currently fit a model and provide components related to the global trend and yearly/weekly seasonality.
• 0:17:43 Live Book Demonstration: A live demonstration using synthetic data showcases Sser's ability to fit a model and predict future values.
• 0:20:41 Model Fitting and Prediction: The speaker demonstrates how to fit a model using Sser and generate predictions, visualizing the results with Vl.
• 0:22:29 Neural Network Visualization: Sser allows visualizing the underlying neural network, showing the components and their interactions.
• 0:25:03 Exploring Sser's Code: The speaker dives into Sser's code, highlighting the preprocessor (adding Fourier terms) and the model architecture.
• 0:27:19 Axon's Role: Axon, an Elixir library for building neural networks, is praised for its descriptive and higher-level abstraction compared to PyTorch.
• 0:28:48 Real Data Example: A real-world example using daily energy pricing from Spain demonstrates Sser's capabilities on a more complex dataset.
• 0:30:16 Future Improvements: The speaker outlines planned improvements for Sser, including handling missing data and incorporating more Neuro-Prophet features.
• 0:30:49 Call to Action: The audience is encouraged to star Sser's GitHub repository and contribute to the "Awesome Machine Learning Elixir" project.