Online Change-Point Detection for Time Series

ISEF Category: Systems Software

Subcategory: Algorithms  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A market can flip in seconds, and a weather station can do the same when a storm front moves in. Your job is to catch that shift before the old pattern stops working. Online change-point detection tries to spot the moment a stream changes behavior. That makes it a great fit for crypto data, weather data, and other live signals.

What Is It?

A change-point is the moment when a data stream stops following one pattern and starts following another. Think of it like listening to a song and noticing the beat suddenly changes. Before the change, the data may rise, fall, or bounce around in a steady way. After the change, the rules seem different.

Online change-point detection means you do this while the data is arriving, not after the full dataset is done. Multivariate time series means you watch several related signals at once, such as price, volume, and spread for crypto, or temperature, pressure, and wind for weather. A parameter-free method tries to avoid hand-tuned thresholds, so the algorithm can adapt without you guessing the right cutoff ahead of time. Provable false-alarm bounds mean the math gives you a limit on how often the method should trigger when nothing has really changed.

Why This Is a Good Topic

This topic works well for a science fair because you can test it on public data, measure performance, and compare it against simpler detectors. You do not need a wet lab, but you do need careful coding and thoughtful evaluation. The real-world link is strong, since fast alerts matter in trading, forecasting, and monitoring. You can learn signal processing, statistics, and algorithm design in one project.

Research Questions

• How does a parameter-free online change-point detector compare with a fixed-threshold detector on public weather-station data?
• What is the effect of using more variables in the multivariate stream on change-point detection accuracy?
• Does the detector find regime shifts faster in crypto order-book data than in weather-station data?
• To what extent do false alarms increase when the input stream becomes noisier or more volatile?
• Which feature set, price and volume or price and spread, gives the clearest change signals in crypto data?
• How does the choice of evaluation window affect precision, recall, and detection delay?

Basic Materials

• Laptop or desktop computer with Python installed.
• Public time series datasets from NOAA, NASA, or a crypto exchange API.
• Text editor or Jupyter Notebook for writing and running code.
• Spreadsheet software for tracking results and plotting summary tables.
• GitHub account or local folder system for version control and file backups.
• Basic statistics reference for precision, recall, and false-alarm rate.

Advanced Materials

• Laptop or workstation with Python and scientific libraries.
• Public API access or bulk downloads for crypto order-book snapshots and weather-station records.
• High-capacity storage for large time series files.
• Jupyter Notebook or a code editor with debugging tools.
• Optional cloud compute for repeated runs across many stations or trading pairs.
• Statistical testing tools for significance tests and confidence intervals.

Software & Tools

• Python: Runs the detector, cleans the data, and automates repeated experiments.
• Jupyter Notebook: Lets you test ideas, plot signals, and document results in one place.
• pandas: Organizes multivariate time series and helps you align timestamps.
• NumPy: Handles array math for fast sliding-window or stream calculations.
• matplotlib: Plots detected change points, scores, and comparison charts.

Experiment Steps

1. Define the stream you will study, then choose one weather dataset, one crypto dataset, or both.
2. Decide what counts as a true change, and build a labeling rule from public events, known weather shifts, or clear signal jumps.
3. Pick baseline detectors to compare against, such as a threshold rule or a windowed mean shift test.
4. Design your feature set, and decide whether you will use raw variables, differences, ratios, or rolling summaries.
5. Build an evaluation plan that measures detection delay, false alarms, and missed changes on the same data splits.
6. Check whether the method stays stable across different sensors, assets, or time periods, then summarize where it works best.

Common Pitfalls

• Treating every spike as a change point, which inflates false alarms on noisy streams.
• Comparing methods on different datasets, which makes the results unfair.
• Using future data when scoring an online detector, which hides real detection delay.
• Choosing change labels by eye without a rule, which makes the ground truth inconsistent.
• Ignoring missing timestamps or irregular sampling, which can break multivariate alignment.

What Makes This Competitive

A stronger project will do more than report that the algorithm finds changes. You can raise the level by testing several datasets, including hard cases with noise, missing points, or weak shifts. You can also compare detection delay and false alarms with a careful statistical test, not just one summary chart. A novel angle, like contrasting crypto market microstructure with weather dynamics, can make the work feel fresh and well thought out.

Project Variations

• Test the detector on NOAA weather-station streams with storms, cold fronts, and sensor drift as the target shifts.
• Apply the same method to crypto order-book depth, spread, and volume to see whether market microstructure changes are easier to catch than weather shifts.
• Compare raw signals with engineered features, such as rolling means and volatility, to see which representation improves online detection.

Learn More

• PubMed review articles on change-point detection: Search for review papers on sequential change detection and online anomaly detection for a research overview.
• NOAA National Centers for Environmental Information: Find station-level weather data, metadata, and climate records for real-world time series.
• NASA Earthdata: Access environmental time series and tutorials on working with satellite and sensor data.
• USGS Earthquake Hazards Program: Use event streams and monitoring data as another example of real-time change detection.
• MIT OpenCourseWare, Introduction to Algorithms: Review algorithm design, complexity, and data-stream thinking in a free course.
• arXiv and peer-reviewed journals: Search for recent papers on online change-point detection, false-alarm bounds, and multivariate streams.