Welcome to a comprehensive guide on Time-Series Analysis, Forecasting, and Machine Learning using Python.
This repository is designed to take you from a foundational understanding to advanced practice. Whether you are dealing with classical forecasting (ARIMA, Prophet), framing time series as supervised machine learning (XGBoost, LightGBM), diving into state-of-the-art Deep Learning (LSTMs, CNNs, Transformers), or tackling advanced industrial use-cases like Anomaly Detection, Clustering, and Classification—this repository provides practical, code-first implementations.
The contents are structured logically: starting with foundational Exploratory Data Analysis (EDA) and statistical analysis, moving through classical methodologies, and transitioning into cutting-edge machine learning and AutoML frameworks.
Not sure which model to use for your forecasting task? Here is a high-level comparison of the approaches covered in this repository to help you decide:
| Approach / Model | Complexity | Best Suited For | Pros | Cons |
|---|---|---|---|---|
| Statistical (ARIMA/SARIMAX) | Low | Small datasets, clear trends/seasonalities. | Highly interpretable, strong mathematical foundation. | Fails on complex non-linear patterns; requires strict stationarity. |
| Tree-Based ML (XGBoost/LGBM) | Medium | Tabularized time series with external features. | Fast, highly accurate, handles non-linearities well. | Requires manual feature engineering (lags, rolling stats). |
| Deep Learning (LSTMs/CNNs) | High | Large datasets, complex sequential patterns. | Learns temporal dependencies automatically. | Computationally expensive; "black box" nature. |
| Transformers | Very High | Massive datasets, long-term forecasting. | State-of-the-art accuracy; captures long-range context. | Extremely data-hungry; prone to overfitting on small data. |
| FBProphet | Low | Business time series (holidays, daily/weekly seasonality). | Works out-of-the-box, handles missing data gracefully. | Less flexible for highly irregular or high-frequency data. |
| AutoML (FLAML) | Low | Robust baselines without manual tuning. | Automatically finds the best algorithm and hyperparameters. | Computationally heavy during the search phase. |
Build a strong foundation in time-series concepts, visualization, and statistical analysis.
- 📋 Datasets Info
- 📚 Introduction to Time Series Analysis (Theory)
- 📊 Time Series Data Visualization
- 🔬 Time Series EDA
- 📉 Time Series Data Analysis
Forecast future values using both classical statistical techniques and modern machine learning.
- 📈 Time Series Forecasting Classical Methods
- ⚙️ Time Series Feature Engineering & Machine Learning
- 🎯 Time Series Forecasting with FBProphet
- 🦾 AutoML For Time Series Forecasting (FLAML)
Leverage neural networks to capture complex, non-linear relationships in sequential data.
- 🕸️ MLPs for Time Series Forecasting
- 🔁 LSTMs for Time Series Forecasting
- 🖼️ CNNs for Time Series Forecasting
- 🤖 Transformers for Time Series Forecasting
Go beyond forecasting to tackle real-world industrial problems.
Created and maintained by Ajit Kumar Singh.