TimeSeries ML
This site is organised around one concrete problem rather than a curriculum. The problem is household energy: given hourly meter readings, forecast consumption, detect anomalies, and work out which appliances are running from the total draw alone. Phisaver Project states that brief, and everything else here is a candidate tool evaluated against it.
That shapes what you will find. The pages are library walkthroughs, mostly following each project’s own tutorials, chosen because a forecasting problem is usually solved by picking the right existing model rather than writing a new one.
Start Here
| Page | Covers |
|---|---|
| Phisaver Project | The brief. What is being predicted (solar, total consumption, aircon) and from which inputs, the three resolutions the meter data arrives at (hourly energy, five-minute power, one-minute sensor readings), and a taxonomy of the model families worth trying: auto-forecast, exponential smoothing, benchmarks, intermittent-demand, multiple-seasonality, and Theta |
Gradient Boosting and Prophet
The pragmatic first attempts. Tree ensembles on lag features are often hard to beat, and Prophet is the standard quick baseline.
| Page | Covers |
|---|---|
| XGBoost | The algorithm itself, worked through from first principles |
| XGBoost: Forecast Energy Consumption | The energy problem end to end: feature engineering from timestamps, training under cross validation, then forecasting forward |
| Prophet | Hourly energy forecasting with Facebook’s Prophet |
Darts
A single API over classical, machine learning, and deep forecasting models, which makes it the easiest way to compare families without rewriting the data pipeline each time.
| Page | Covers |
|---|---|
| Darts | The long tour: training and predicting, inspecting seasonality, hyper-parameter search with Theta, backtesting against history, global models trained across many series, N-BEATS for the deep learning case, covariates and encoders for external data, and probabilistic forecasts with their distributions |
| Hyper-parameter Optimization | Tuning a deep model for electricity load with confidence intervals: a linear baseline, then TCN, then Optuna over its parameters, then picking the winner |
Nixtla
The Nixtla stack splits by model type across four libraries, so the pages are grouped the same way. Most follow the upstream tutorials.
| Page | Covers |
|---|---|
| StatsForecast | Training many statistical models over many series at once, evaluating them, and selecting the best per series |
| MLForecast | The machine learning arm, including distributed training via Spark and LightGBM |
| NeuralForecast | Fitting LSTM and NHITS on AirPassengers, then plotting the predictions |
| Exogenous Variables | Training and forecasting with external inputs, both historical and known-future |
| Save and Load Models | Persisting a trained neural forecaster and reading it back |
| Neural Advanced | The multi-series version: train several neural models, evaluate, and select per series |
| All Methods | The comparison that ties the stack together. Statistical, machine learning, and neural forecasts over the M5 dataset, cross-validated and evaluated per series and per window |
| Hierarchical Forecast | Placeholder. A title and a link, no content yet |
The FFT forecasting, electricity load and demand peak pages moved to Electrical Docs.
Signal Analysis and Explainability
| Page | Covers |
|---|---|
| FFT Analysis | Building a synthetic multi-frequency signal, then using wavelets and FFT features to feed a random forest regressor |
| Xplainable | The xplainable library on the Titanic dataset, using a partitioned classifier |
| Xclassifier | XClassifier on the same data, closely overlapping the previous page |
Not Covered Yet
Stated plainly, so the gaps are not mistaken for oversights:
- The project brief is not answered here. Phisaver Project sets out consumption forecasting, anomaly detection, and appliance disaggregation. The pages that follow evaluate tools; none of them close the loop on the actual data.
- Nothing on disaggregation. Recovering per-appliance draw from a total signal is the hardest of the three goals and has no page at all.
- No anomaly detection, despite it being in the brief.
- Three stubs. Financial Analysis is a single unticked to-do about the US Treasury yield curve, Hierarchical Forecast is a bare link, and the two
xplainablepages duplicate each other. - sktime is named as a candidate framework in the brief but never evaluated.
- Most pages follow upstream tutorial data rather than the project’s own meter readings, so the comparisons are indicative rather than decisive.