Predicting favorable and dangerous weather conditions across Europe using supervised machine learning — built for a European climate nonprofit.
ClimateWins is a European nonprofit concerned with the increasing frequency of extreme weather events over the past two decades. This project delivers a supervised machine learning pipeline that predicts whether daily weather conditions at a given station will be favorable or unfavorable — with the longer-term goal of predicting dangerous conditions in advance.
This is a complete, end-to-end data science project covering business framing, exploratory data analysis, feature engineering, model development, business recommendations, stakeholder presentation, and ethical analysis.
ClimateWins needs a reliable, explainable model that:
- Predicts favorable vs unfavorable weather conditions from daily measurements
- Minimises false positives — dangerous days incorrectly labelled as safe
- Is interpretable enough for non-technical stakeholders to trust and act on
- Generalises to genuinely unseen future dates, not just held-out random samples
Success definition: A model meaningfully above the 76.5% majority-class baseline, with particular focus on reducing false reassurance (false positives) — the highest-cost error type for organisations making operational decisions based on weather predictions.
| Property | Detail |
|---|---|
| Source | European Climate Assessment & Dataset (ECA&D) |
| Stations | 18 weather stations across mainland Europe |
| Period | January 2000 – January 2010 (3,653 daily records) |
| Features | Temperature, global radiation, cloud cover, humidity, precipitation, sunshine, wind, pressure |
| Target | Binary: favorable / unfavorable day (based on picnic weather proxy label) |
| Missing values | None |
Key dataset characteristics identified during EDA:
- Sonnblick (alpine, 3,100m) excluded — 0% favorable days across entire period, unsuitable for binary classification
- Strong temporal autocorrelation — consecutive days are related, requiring chronological splitting
- Class imbalance: 74.3% unfavorable, 25.7% favorable at Basel reference station
- Seasonal patterns confirmed coherent — data passes physical sanity checks
ClimateWins/
├── Data/
│ ├── Supervised/
│ │ ├── weather_prediction_dataset.csv # Raw weather features
│ │ ├── weather_prediction_picnic_labels.csv # Raw target labels
│ │ └── weather_merged.csv # Cleaned, merged dataset
│ └── Unsupervised/
├── Notebooks/
│ ├── 01_data_understanding.ipynb # Loading, merging, cleaning
│ └── 02_supervised_learning_basel.ipynb # EDA, feature analysis, modelling
├── Deliverables/
│ ├── ClimateWins_Presentation.pptx # Stakeholder presentation
│ └── Exercise_1_2_Ethics_and_Data_Preparation.docx # Ethics and data write-up
└── README.md
- Merged weather observations with picnic labels on DATE key
- Converted DATE from integer format (YYYYMMDD) to datetime
- Removed incomplete 2010 record (1 day — statistically meaningless)
- Verified zero missing values across all 3,653 records
- Confirmed seasonal temperature coherence (northern hemisphere pattern)
- Identified Sonnblick as a distributional outlier — excluded from modelling
- Analysed class balance across all 18 stations (range: 0% to 48.5% favorable)
- Constructed feature correlation matrix to identify multicollinearity
Starting from 9 Basel weather features, reduced to 5 based on:
- Target correlation analysis — global radiation (0.656) and sunshine (0.613) are strongest predictors
- Multicollinearity removal — temp_min, temp_max, and sunshine dropped (correlation >0.85 with retained features)
- Low-signal removal — pressure dropped (target correlation: 0.041)
Final features selected:
| Feature | Target Correlation | Rationale |
|---|---|---|
BASEL_global_radiation |
+0.656 | Strongest predictor; proxy for sunshine signal |
BASEL_temp_mean |
+0.562 | Representative temperature; min/max redundant |
BASEL_cloud_cover |
-0.424 | Independent negative signal |
BASEL_humidity |
-0.443 | Independent negative signal |
BASEL_precipitation |
-0.257 | Independent negative signal |
Critical decision: chronological split, not random shuffle.
Weather data has strong temporal autocorrelation. A random split allows the model to see dates adjacent to test dates during training — inflating accuracy artificially. The chronological split ensures the model only ever learns from the past to predict the future.
| Split | Period | Records | Favorable Rate |
|---|---|---|---|
| Train | 2000–2007 | 2,922 | 26.2% |
| Test | 2008–2009 | 731 | 23.5% |
Seasonal distribution validated as consistent between splits — test results reflect genuine generalisation.
StandardScaler applied — fitted on training data only, applied to both sets. Prevents test set information leaking into preprocessing.
| Model | Test Accuracy | False Positives | False Negatives | Recommended |
|---|---|---|---|---|
| Baseline (majority class) | 76.5% | — | — | No |
| KNN (k=5) | 92.2% | 45 | 12 | Fallback only |
| Decision Tree (depth=4) | 98.2% | 8 | 5 | Yes |
Best model: Decision Tree (max_depth=4)
- 21.8 percentage points above baseline
- 82% reduction in dangerous false positives vs KNN
- Train accuracy 97.8% / Test accuracy 98.2% — gap of -0.4% confirms no overfitting
- Root node splits on global radiation — consistent with feature correlation analysis
- Fully interpretable — decision logic is visible and explainable to stakeholders
1. Global radiation is the dominant predictor.
The Decision Tree's root split uses global radiation at threshold 0.657. This single feature separates the majority of favorable from unfavorable days — consistent with its 0.656 correlation with the target. It functions as a composite signal encoding sunshine, cloud cover, and season simultaneously.
2. The chronological split was the most important methodological decision.
A random split would have produced artificially inflated accuracy due to temporal leakage. The chronological split ensures results reflect genuine forecasting ability on unseen future dates.
3. Accuracy alone is a misleading metric for this problem.
The baseline model achieves 76.5% accuracy by predicting "unfavorable" every day — learning nothing. Every model must be evaluated against this baseline and assessed on false positive rate specifically, given the asymmetric cost of dangerous false reassurance.
4. Station-level variation is significant.
Favorable day rates range from 0% (Sonnblick) to 48.5% (Perpignan). A single global model cannot serve all stations equally. Future work should evaluate station-specific models for locations with atypical climate profiles.
Recommendation 1 — Deploy the Decision Tree model for Basel. With 98.2% accuracy and only 8 false positives across a 2-year test period, the model is reliable enough to inform operational decisions. Its interpretability — the decision path is fully visible — allows stakeholders to understand and trust each prediction.
Recommendation 2 — Prioritise false positive reduction over accuracy. For high-stakes use cases (agricultural planning, infrastructure management), consider raising the classification threshold above 50% to further reduce dangerous false reassurance, accepting more unnecessary caution in exchange.
Recommendation 3 — Expand to additional stations with validated models. Basel serves as the proof-of-concept. The same methodology should be applied to each station individually, with Sonnblick excluded and stations with very low favorable rates (Oslo: 16.9%) treated with additional caution.
Recommendation 4 — Extend the historical window. The 2000–2009 dataset is too short to detect climate change trends reliably. Incorporating ECA&D data from the late 1800s would enable long-term trend analysis and strengthen the model's exposure to historical extreme events.
A full ethical analysis is documented in Deliverables/Exercise_1_2_Ethics_and_Data_Preparation.docx. Key concerns are summarised below.
Proxy label risk The "favorable day" label is based on a picnic weather definition — a human judgement call. The model is only as meaningful as this definition. For safety-critical applications, the label must be reviewed and validated by domain experts before deployment.
Temporal bias The model is trained on 2000–2007 data. Climate change — the very phenomenon ClimateWins is concerned about — means historical patterns may no longer represent future conditions accurately. Regular retraining on more recent data is essential.
Geographic bias Results are validated for Basel only. Applying this model to other stations without station-specific revalidation risks unreliable predictions in different climate regimes (alpine, Mediterranean, Scandinavian).
Automation and human oversight Model predictions should inform decisions, not replace them. For safety-critical warnings, human review of borderline predictions is required. The Decision Tree's full interpretability — its decision path is visible and auditable — directly supports this governance requirement.
Explainability as an ethical requirement A model that cannot explain its predictions is difficult to challenge or audit. The Decision Tree was selected partly for this reason: stakeholders can see exactly which weather thresholds drive each prediction.
File: Deliverables/ClimateWins_Presentation.pptx
A 10-slide stakeholder-facing presentation designed for a non-technical business audience. Covers the business problem, data foundation, key EDA findings, methodology, model performance results, error type analysis, recommendations, ethical considerations, and a closing summary. No code or technical jargon — built to communicate findings to ClimateWins leadership and partner organisations.
| Slide | Content |
|---|---|
| 1 | Title and project framing |
| 2 | The problem ClimateWins is solving |
| 3 | Data foundation — key statistics |
| 4 | What the data revealed — 4 EDA findings |
| 5 | How the prediction system was built — 4-step methodology |
| 6 | Model performance results with bar chart |
| 7 | Why error type matters more than accuracy |
| 8 | Four prioritised business recommendations |
| 9 | Ethical considerations |
| 10 | Closing summary with headline statistics |
File: Deliverables/Exercise_1_2_Ethics_and_Data_Preparation.docx
A formal written deliverable covering all four Exercise 1.2 requirements:
-
Ethical issues in machine learning and automation — why ethical issues arise, the amplification effect of automated decisions, and five specific concerns for this project including label definition risk, temporal bias, geographic limitation, automation governance, and explainability.
-
Data landscape — source and provenance (ECA&D), full dataset structure, weather variables measured, and three key landscape findings: station-level variation, Sonnblick exclusion rationale, and temporal data structure.
-
Supervised vs unsupervised learning — definitions, application to ClimateWins, and a comparison table covering training data, goal, output type, evaluation approach, and project-specific use case for each paradigm.
-
Data preparation — documented decisions covering loading and merging, quality assessment, feature selection with justification table, chronological train/test split rationale, and feature scaling protocol including the critical distinction between fitting and transforming.
- Python 3.x
pandas— data manipulation and analysisscikit-learn— preprocessing, modelling, evaluationmatplotlib/seaborn— visualisationnumpy— numerical operations
Benjamin Akingbade Data Analyst | Python · scikit-learn · PostgreSQL · FastAPI LinkedIn • Portfolio