1. General Policy

The project must be completed in pairs (Any exception requires instructor approval).

Originality in dataset selection, forecasting horizon definition, and modeling strategy will be positively evaluated.

2. Project Tracks

Students must choose one of the following tracks:

Track 1 Research Replication & Improvement

Goal: Select a peer-reviewed research paper (or high-quality technical study) on time series, reproduce it, and attempt to improve it.

• Mandatory: Reproduce the original results or core methodology using the original (or similar) data.
• Improvement: Attempt a meaningful improvement.
  • Applying the method to a new domain/dataset.
  • Tuning hyperparameters or using a different optimization method.
  • Combining with another model (Hybrid approach).
  • Improving the evaluation metric or forecast horizon.
• Focus: Scientific rigor and critical analysis of the original paper's claims.

Track 2 Educational Method Deep-Dive

Goal: Independent study of an advanced forecasting method not covered in class.

• Scope: Examples include GARCH, Prophet, N-BEATS, Transformer for Time Series, Bayesian Structural Time Series, or Vector Autoregression (VAR).
• Mandatory Requirements:
  • Theory: Clear explanation of the mathematical/algorithmic foundations.
  • Implementation: Working Python code (not just using a library blindly—explain the parameters).
  • Analysis: Apply the model to real data and compare it against a standard baseline (e.g., SARIMA).
• Warning: Purely theoretical slides without code, data analysis, and model diagnostics will receive a low grade.

Track 3 Applied Forecasting Project

Full real-world dataset project with end-to-end responsibility.

• Mandatory: Data cleaning, handling missing values, seasonality detection, comparison of at least two models, forecast horizon definition, and rigorous evaluation (MAE, RMSE, MAPE).
• Encouraged: Financial data, energy consumption, epidemiology, transportation, macroeconomic data, or self-collected data.
• Focus: Uncertainty quantification and actionable insights.

3. Technical Requirements (Mandatory for All)

Each project must include:

• Time plot: Clear visualization of the raw series.
• Decomposition: Trend, seasonal, and residual components.
• Stationarity test: ADF (Augmented Dickey-Fuller) at minimum.
• ACF and PACF plots: Before and after differencing.
• Residual diagnostics: White noise tests, Ljung-Box test, normality checks.
• Forecasting: At least one forecast with confidence intervals.
• Comparison: At least two models must be compared.

Students must interpret coefficients (AR/MA terms) and explain their physical meaning where possible.

4. Submission Format

Required Items:

• Link: Google Colab (fully executable) or GitHub repository.
• Email Submission: ZIP archive containing the final report (PDF, max 5 pages), slides, dataset description, and AI Usage Disclosure section.

Projects must be fully reproducible.

5. AI Usage Policy (Mandatory Disclosure)

AI tools may be used only as support tools. Every project must include a section named "AI Usage Disclosure". This disclosure must appear in ALL of the following places:

• Report: dedicated section
• Notebook/Colab: first markdown cell titled "AI Usage Disclosure"
• Slides: final slide titled "AI Usage Disclosure"
• Repository/ZIP: AI_USAGE.md file

Students must explicitly state:

• Which AI tools were used.
• For what purpose and to what extent.
• What was rewritten or modified manually.
• What was independently verified.

6. Reflection & Technical Ownership (Mandatory)

Each report must include a Personal Reflection and Technical Ownership section (Min. one full page).

Students must describe:

• What they personally implemented.
• What failed (e.g., model failed to converge, residuals were not white noise) and how they fixed it.
• Technical challenges encountered.
• What they learned and what surprised them about the data.
• What they would improve.
• Failure log: 2–3 concrete modeling attempts that failed.
• Team contribution log: What each teammate did; short work log.

This section has significant grading weight.

7. Ethical Considerations (Mandatory)

Each project must include an Ethical and Responsible Forecasting Statement addressing:

• Misuse of forecasts: How could a wrong forecast harm stakeholders?
• Economic consequences: Impact of errors on decision making.
• Overfitting risks: False confidence in model precision.
• False confidence intervals: Underestimating risk/uncertainty.
• Data quality: Issues with missing data or outliers.
• Responsible interpretation: Transparency about model limitations.

Projects that ignore the consequences of poor forecasting will receive lower evaluation.

8. Oral Technical Mini-Defense

After the presentation, each team will undergo a 5-minute technical Q&A.

Possible questions:

• Why did you choose differencing order d=1?
• Is the series stationary? How do you know?
• Why did you include/exclude seasonality?
• What does the Ljung-Box test tell you about your residuals?
• Explain the meaning of the AR(1) coefficient in your model.

Performance in the mini-defense affects the Reflection score. This ensures authentic understanding and technical ownership.

9. Timeline

• Week 4: Project proposal (max 2 pages). Includes track selection, dataset description, initial visualization, and planned methods. Instructor approval required.
• Week 5–6: Feedback and scope refinement.
• Final week of the semester: Presentation (20–30 minutes, 20–40 slides). Exact dates on Moodle.
• Final Submission: Up to 4 weeks after presentation. No late submissions without approval.
• Calendar dates: Exact deadlines will be announced on Moodle.

10. Grading Rubric (50% of Course Grade)

This rubric sums to 100% and is applied to the project grade (which constitutes 50% of the course grade).

11. Academic Integrity

The following are prohibited:

• Full AI-generated projects.
• Undocumented AI assistance.
• Copy-paste tutorials.
• Using "Auto-ARIMA" or auto-forecasting tools without full explanation and manual validation.
• Evaluating models on training data.

12. Bonus (Up to +5%)

Awarded for highly original datasets, multivariate analysis (VAR/Exogenous variables), complex seasonality handling, or deep uncertainty quantification.