Week 15 Final Project Report Guide - Time Series Analysis

Overview

The final project is assessed primarily through a written report (PDF). The evaluation emphasizes publication readiness — your work should meet approximately 70% of the requirements of a scientific paper (IMRAD structure, labeled figures, reproducible methods, honest discussion).

Deliverables (submit before or at class)

Final project report (PDF) — extended and improved from the midterm baseline
All notebooks / code (GitHub link or Kaggle notebook)

Recommended Report Organization

1) Introduction: Background → Challenges → Motivation

Background: State the problem clearly and explain why this project/research is needed in this field.
Challenges: Explain why related or comparison methods are not fully sufficient, or what limitations they have in this specific problem context.
Motivation for proposed methodology: Show how those gaps motivate your analytical workflow or model design.
Research questions: End the section with 3–5 research questions aligned with your proposed approach.

Contribution check: The introduction should not only say "we forecast X"; it should explain what gap your method addresses and why the result matters.

2) Related Work (with Citations)

Introduce key related/comparison methods used in the same or similar problem domain.
Cite sources properly and summarize their strengths and limitations in your context.
Explain which baseline/comparison methods you will use in your report and why.
If you do not include a standalone Related Work section, integrate related/comparison methods with citations into the Introduction.

3) Methodology: Dataset/Cohort + Proposed Method

Dataset / cohort / entity description: data source, sample size, time range, sampling frequency, unit of analysis, target variable, and key covariates.
Data preparation protocol: missing data handling, resampling, stationarity transformation, scaling, windowing, feature engineering, and leakage prevention.
Proposed method: clearly describe your main method or pipeline. Do not only list models; identify which method is your proposed approach.
Comparison design: specify baselines (e.g., naive, ARIMA/SARIMA, Random Forest/XGBoost, LSTM/GRU/TCN) and explain why they are appropriate.
Reproducibility: report train/test split, validation strategy, hyperparameters, random seeds, software libraries, and source code link.

4) Results and Evidence

Present numbered tables/figures with captions and clear axis labels.
Report relevant metrics such as MAE, RMSE, MAPE, accuracy, F1-score, AUC, or domain-specific metrics.
Compare proposed method against baselines and explain why differences occur.
Include error analysis: where does the model fail, and what temporal patterns are missed?

5) Discussion, Limitations, and Conclusion

Interpret findings in the context of stakeholder decisions.
Translate results into practical recommendations.
State limitations honestly: dataset bias, short history, missing variables, unstable seasonality, overfitting, or deployment constraints.
Conclude with the main contribution and concrete future improvements.

Title Writing Guide

Use this pattern:
<Method> + <Task> + <Objective or Dataset/Target Entity>

Question 1: What is the main method?
Question 2: What task does it serve (forecasting, classification, anomaly detection, etc.)?
Question 3: Who is the target entity or audience receiving value from the data (e.g., lung cancer patient, hospital, grid operator, call center manager)?

Example Title

LSTM-based Time Series Forecasting for Emergency Healthcare Call Volume Prediction in Public Hospital Systems

(Method = LSTM, Task = forecasting, Objective/Target Entity = emergency healthcare calls in public hospital systems)

Contribution-Driven Writing (Most Important)

Your report should be built around a clear contribution. The contribution appears first in the Introduction, is justified in Related Work, implemented in Methodology, supported in Results, and restated in the Conclusion.

Contribution Logic

Limitation in related/comparison methods → Gap → Your contribution → Evidence

Limitation: What do existing or baseline methods fail to handle well in your context?
Gap: What specific forecasting/classification need remains unmet?
Contribution: What do you introduce (model, feature pipeline, hybrid workflow, evaluation design, interpretability analysis)?
Evidence: What result proves the contribution (better metric, clearer error behavior, stronger interpretability, more useful decision support)?

Contribution Statement Template

“Existing approaches [limitation]. To address this gap, we propose [your method/workflow], which enables [practical objective]. Compared with [baseline], our approach achieves [evidence: metric/insight/actionability].”

How to Repeat the Contribution

Introduction: state the contribution claim explicitly after motivation.
Related Work: position the contribution against limitations of prior methods.
Methodology: show where the contribution appears in the pipeline/design.
Results: provide quantitative or visual evidence for the contribution claim.
Discussion: explain why the contribution matters for stakeholders and decisions.
Conclusion: restate the contribution as the main takeaway.

Quality check: If a reader cannot identify your contribution in 2–3 sentences, the report is still too descriptive.

Sample Papers to Read Before Writing

Prioritize method papers (not surveys) so you can learn how authors present a concrete model, experimental design, and evidence. You do not need to copy the methods exactly; use them as writing and structure references.

N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting — useful for writing model motivation, architecture design, and benchmark comparison sections. Read paper
DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks — useful for probabilistic forecasting setup, likelihood-based training, and uncertainty reporting. Read paper
Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting — useful for multi-horizon design, feature importance interpretation, and ablation reporting. Read paper
InceptionTime: Finding AlexNet for Time Series Classification — useful for classification projects and strong baseline-comparison methodology. Read paper
Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting — useful for long-horizon forecasting papers and efficient architecture justification. Read paper
Forecasting at Scale (Prophet) — useful for classical + practical forecasting structure, trend/seasonality decomposition, and business-oriented discussion. Read paper
A Comparison of ARIMA and LSTM in Forecasting Time Series — useful as a direct example of classical-vs-deep-learning experimental comparison. Read paper

Reading task: For each paper you use, extract (1) contribution statement, (2) comparison baseline, (3) evidence table/figure design, and (4) limitation discussion style.

Report Checklist

Start with a clear problem statement section.
Write the Introduction as Background → Challenges → Motivation → Research Questions.
Include related/comparison methods with citations and explain their limitations.
State the main contribution explicitly and repeat it in Results and Conclusion.
Label all figures clearly: axis names, units, legends, metric definitions.
Show a model comparison table (not just individual results).
Include at least one ablation or insight (e.g., "removing lag features drops MAE by 30%").
Methods must be reproducible: mention hyperparameters, validation setup, random seeds.
Mandatory: Include a source code link in the report (GitHub / Kaggle).
End with takeaway bullets: what works, what fails, what is next.

Evaluation Rubric (10-Point Scale)

Your final project is graded on 5 criteria, 2 points each. Each criterion is evaluated against both course standards and publication readiness.

Criterion 1: Problem, Gap & Contribution

Score	Description
2.0	Introduction clearly follows Background → Challenges → Motivation, identifies limitations of related/comparison methods, and states a specific contribution plus 3–5 aligned research questions.
1.5	Problem is well-motivated and contribution is present, but related-method limitations or research questions are not sharply articulated.
1.0	Problem stated but mostly descriptive ("we forecast X"). Gap/contribution is vague or not linked to related methods.
0.5	Vague or missing problem statement. Dataset introduced without context.

Publication check: Could a reviewer understand why this problem matters from the introduction alone?

Criterion 2: Related Work, Methodology & Technical Depth

Score	Description
2.0	Related/comparison methods are cited and summarized. Methodology clearly describes dataset/cohort, preprocessing protocol, proposed method, baselines, validation, hyperparameters, and reproducibility details.
1.5	Methodology is mostly complete but has gaps (e.g., related work too brief, baseline choice weakly justified, temporal split not fully explained).
1.0	Methods are listed but the proposed method is unclear, citations are weak/missing, or preprocessing/validation has limited justification.
0.5	Single model with no pipeline structure. Significant methodological errors (data leakage, wrong evaluation).

Publication check: Is the methodology reproducible? Could someone replicate this from the description?

Criterion 3: Results, Comparison & Analysis

Score	Description
2.0	Clear comparison table (MAE/RMSE across all models). Forecast overlay plots. Honest discussion of why one model outperforms another. At least one ablation or insight. Limitations and future work explicitly stated.
1.5	Comparison table present but analysis is shallow ("XGBoost is best" without explaining why). Plots present but not well-discussed. Limitations mentioned briefly.
1.0	Results reported for individual models but no structured comparison. Missing metrics or forecast plots. No discussion section.
0.5	Minimal or incorrect results. No comparison across methods.

Publication check: Does the results section tell a story, or just list numbers?

Criterion 4: Report Quality & Scientific Writing

Score	Description
2.0	Report follows clear IMRAD structure (Introduction → Data → Method → Results → Discussion → Conclusion). Figures labeled with captions, axes, legends. Tables formatted consistently. Writing is concise. Code submitted and organized. References cited where appropriate.
1.5	Structure present but some sections weak (e.g., missing conclusion, unlabeled figures, inconsistent formatting). Code submitted but not well-organized.
1.0	Report reads like a notebook dump — code outputs pasted without narrative. No clear section structure. Figures without labels.
0.5	Report missing or extremely incomplete. No code submitted.

Publication check: If you removed the code cells, would it read like a coherent draft paper?

Criterion 5: Report Coherence & Communication

Score	Description
2.0	Clear report narrative with logical flow from problem to conclusion. Writing is concise and evidence-based. Key decisions are justified. Source code link included.
1.5	Most sections are present, but flow is choppy or explanations are too descriptive. Some claims are not fully supported by evidence.
1.0	Report is disorganized or incomplete in key parts. Technical choices are listed but weakly justified.
0.5	Missing critical sections, or report does not clearly explain methodology and findings.

Publication check: Could this report draft be submitted to a student workshop with minor revisions?

Score Summary

#	Criterion	Max	Publication Parallel
1	Problem, Gap & Contribution	/2	Introduction
2	Related Work, Methodology & Technical Depth	/2	Related Work + Methodology
3	Results, Comparison & Analysis	/2	Results + Discussion
4	Report Quality & Scientific Writing	/2	Overall paper quality
5	Report Coherence & Communication	/2	Scholarly communication
	Total	/10

How the Final Differs from Midterm

	Midterm (30%)	Final (50%)
Models	ARIMA/SARIMA baseline only	Classical + ML + DL (DL encouraged)
Report	Baseline report	Full IMRAD-style report with explicit proposed method
CLOs	1.2–4.1	1.1–5.2 (all CLOs)
Focus	EDA + classical baseline quality	Integration, comparison, depth, publication readiness

← Back to Schedule View Grading Form