Open Multidomain Time Series

Level: Hard
Goal: Choose a series from a curated collection and design a custom forecasting/classification problem.

Collection

• Source: Open Time Series Datasets – GitHub
• Link: https://github.com/liaoyuhua/open-time-series-datasets

Download & Load Instructions

1. Open the GitHub link above.
2. Find a dataset (e.g., airline passengers, air pollution, traffic).
3. Click file → "Download" or "Raw" and save as CSV to data/open_ts/.

Data Loading (Example)

import pandas as pd

df = pd.read_csv("data/open_ts/example.csv")  # adjust
df["Date"] = pd.to_datetime(df["Date"])       # adapt column name
df = df.set_index("Date").sort_index()

Implementation Steps

1. Dataset Selection

• Browse the open time series collection
• Select a dataset that interests you
• Consider:
  • Data availability and quality
  • Problem complexity
  • Domain knowledge
  • Research questions

2. Problem Formulation

• Define clear objectives:
  • Forecasting (univariate or multivariate)
  • Classification
  • Anomaly detection
  • Pattern recognition
• Formulate research questions
• Define success metrics

3. Data Exploration

• Load and inspect selected dataset
• Understand data structure and variables
• Examine data quality (missing values, outliers)
• Analyze data frequency and time range
• Explore domain-specific characteristics

4. Exploratory Data Analysis (EDA)

• Plot time series
• Identify trends, seasonality, cycles
• Perform time series decomposition
• Calculate and visualize ACF/PACF
• Analyze relationships between variables (if multivariate)
• Apply domain knowledge

5. Data Preprocessing

• Handle missing values appropriately
• Detect and handle outliers
• Apply transformations (log, differencing) if needed
• Test for stationarity
• Prepare data for modeling

6. Model Building

• Classical Methods:
  • ARIMA/SARIMA models
  • Exponential smoothing
• Machine Learning:
  • Feature engineering (lags, rolling stats, calendar features)
  • Tree-based models (Random Forest, XGBoost)
  • Linear models with features
• Deep Learning (optional):
  • LSTM/GRU networks
  • CNN for time series
• Hybrid Approaches:
  • Combine multiple methods
  • Ensemble forecasts

7. Model Evaluation

• Design appropriate train/validation/test splits
• Use time series cross-validation
• Calculate relevant metrics
• Compare multiple approaches
• Analyze model performance

8. Advanced Analysis (Customize Based on Problem)

• If Forecasting:
  • Generate future forecasts
  • Analyze forecast uncertainty
  • Compare forecast horizons
• If Classification:
  • Build classification models
  • Evaluate classification performance
  • Analyze feature importance
• If Anomaly Detection:
  • Detect anomalies
  • Analyze anomaly patterns
  • Validate detections

9. Interpretation and Discussion

• Interpret results in domain context
• Discuss implications
• Identify limitations
• Suggest future work
• Compare with existing literature

Expected Deliverables

1. Problem Definition:

   • Clear problem statement
   • Research questions
   • Success criteria

2. EDA Report:

   • Comprehensive data exploration
   • Visualizations
   • Domain-specific insights

3. Model Results:

   • Multiple model approaches
   • Performance comparison
   • Best model selection
   • Detailed results

4. Code:

   • Complete, well-documented notebook
   • Reusable functions
   • Clear structure

5. Discussion:

   • Interpretation of results
   • Domain insights
   • Limitations and future work

Tips

• Choose a dataset that matches your interests and skill level
• Start with a clear problem definition
• Apply domain knowledge throughout
• Try multiple approaches and compare
• Document all decisions and rationale
• Consider both classical and ML methods
• Use appropriate evaluation metrics for your problem
• Interpret results in domain context
• This is an opportunity to be creative and explore
• Consider real-world applications and implications
• Consult domain literature for context
• Think about what makes your analysis unique or valuable