XGBoost Time Series Outlier Detection

A system for detecting outliers in time series data using XGBoost regression with dynamic thresholding, optimised for handling periodic patterns.

Requirements

• Python 3.13
• uv (for dependency management)
• Node.js/Deno (for synthetic data generation)

Dependencies

See pyproject.toml for more details

Components

1. Data Generator

Run using Deno:

deno run gen_synth_data.js > synthetic_data.csv

Generates time series data with:

• Daily observations (100 points from 2023-01-01)
• Upward trend (+0.5/day)
• Two step changes (+20 at day 30, +15 at day 60)
• Cyclical patterns (10-point amplitude)
• Random noise (±1.5 points)

2. Outlier Detector

from outlier_detection import detect_outliers, plot_time_series

# Load data
df = pd.read_csv("synthetic_data.csv")

# Detect outliers
df_with_outliers = detect_outliers(df)

# Visualize (saves to plots/time_series_outliers_{sheet_name}.html)
fig = plot_time_series(
    df_with_outliers,
    "Time Series Analysis",
    "dataset_name",
    df_with_outliers["is_outlier"]
)
fig.show()

Key Features:

• Temporal feature engineering (day/month/year patterns)
• Dynamic thresholding for periodic spikes
• Interactive Plotly visualizations
• Excel and CSV support