FundFirst — Explainable Deposit Feasibility Classification

An explainable machine-learning prototype that classifies first-time buyer deposit-saving scenarios across six South-West London boroughs as Achievable, Stretch or Unfeasible.

Full Case Study • Modelling Notebook • Evaluation Results • Live Application

FundFirst explores whether deposit-saving scenarios for prospective first-time buyers in six South-West London boroughs can be classified as Achievable, Stretch or Unfeasible.

The project combines public housing, earnings, household-saving and interest-rate data; defines a Transparent Saving Model (TSM)** labelling rule; compares Logistic Regression with Random Forest; and uses SHAP plus an income-stratified performance audit to examine the selected model.

The final Logistic Regression pipeline is served through a FastAPI backend and connected to a Lovable frontend for live predictions.

FundFirst measures agreement with project-generated feasibility labels. It does not predict mortgage approval, observed home purchases or individual affordability, and it is not financial advice.

FundFirst is deployed as an end-to-end machine-learning application with a separate frontend and Python inference API.

Launch FundFirst • View FastAPI Backend • Homepage Screenshot • Results Screenshot

User enters a scenario
        ↓
Lovable frontend
        ↓
HTTPS POST /predict
        ↓
FastAPI backend
        ↓
Saved scikit-learn pipeline
StandardScaler + Logistic Regression
        ↓
Prediction + class probabilities
        ↓
JSON response
        ↓
Lovable displays the result

The frontend handles the user experience, while preprocessing and model inference remain in the Python backend.

The Lovable frontend does not recreate the trained model, scaling logic or prediction rules. The FastAPI service remains the source of truth for live predictions.

Public housing + economic data → transparent feasibility labels → chronological ML evaluation → SHAP explanations → fairness checks

FundFirst classifies deposit-saving scenarios into three feasibility tiers:

• Achievable — up to 72 months
• Stretch — 73–108 months
• Unfeasible — over 108 months

The final dataset contains 72 borough-year observations for Croydon, Kingston upon Thames, Merton, Richmond upon Thames, Sutton and Wandsworth from 2014–2025.

Models were evaluated chronologically:

• 2014–2022 — initial training
• 2023 — validation checkpoint
• 2014–2023 — final model refit
• 2024–2025 — held-out test period

Logistic Regression was selected because it performed better on the held-out period while retaining a simpler and more interpretable model structure.

On the 12-row test set, it correctly classified all Achievable and Stretch observations, but recall for Unfeasible cases was 0.33.

FundFirst evaluated four assurance dimensions:

The income-stratified audit flagged a 33.3 percentage-point accuracy gap and a 66.7 percentage-point Stretch-precision gap between borough groups.

These results are treated as investigation flags rather than evidence of unfair treatment, given the small 12-row test set, aggregate borough-level data and use of income as a socioeconomic proxy.

The strongest assurance evidence came from transparency and reproducibility, while fairness and robustness remained open risks.

View full evaluation results →

1. Clean and combine four official public datasets at borough-year level.
2. Calculate a 10% deposit target and assume monthly saving equal to 20% of gross income.
3. Assign TSM labels:
   • Achievable at ≤72 months
   • Stretch at >72 and ≤108 months
   • Unfeasible at >108 months
4. Compare standardised Logistic Regression with a 300-tree Random Forest using chronological splits.
5. Explain held-out predictions with local and global SHAP values.
6. Audit accuracy and class precision across income-stratified borough groups.

The 72-month Achievable threshold was calibrated using the 2014–2022 development period to retain a usable three-class structure. It is specific to this dataset and is not a universal affordability definition.

The suppressed Kingston upon Thames earnings value for 2018 is linearly interpolated between 2017 and 2019.

The resulting dataset contains aggregate statistics only and no personal records.

• Raw Data Notes →
• Clean Data Notes →

• Labels are generated by a deterministic rule rather than observed buyer outcomes.
• The dataset is small and covers only six South-West London boroughs.
• National saving-ratio and Bank Rate features repeat across boroughs within each year.
• The held-out test contains only 12 observations, so performance and audit gaps are prototype-level evidence.
• SHAP describes model contributions, not causal effects.
• Income is used only as a socioeconomic proxy in the fairness analysis.
• The project is an educational research prototype, not a production decision system.

fund-first/
├── case-study/
│   └── README.md
│
├── data-clean/
│   ├── README.md
│   ├── base_rate_annual_clean.csv
│   ├── earnings_annual_clean.csv
│   ├── fundfirst_features_clean.csv
│   ├── fundfirst_labelled_data.csv
│   ├── hpi_annual_clean.csv
│   └── saving_ratio_annual_clean.csv
│
├── data_raw/
│   ├── README.md
│   ├── earnings.xls
│   ├── hpi.csv
│   ├── rate.csv
│   └── saving_ratio.csv
│
├── logo/
│   └── logo.png
│
├── notebook/
│   ├── README.md
│   └── fundfirst-analysis.ipynb
│
├── outputs/
│   ├── README.md
│   ├── bias_audit_accuracy.csv
│   ├── bias_audit_precision.csv
│   ├── bias_audit_precision_gaps.csv
│   ├── model_test_results.csv
│   └── test_predictions.csv
│
├── results/
│   ├── README.md
│   ├── held_out_model_performance.png
│   ├── lr_test_confusion_matrix.png
│   ├── rf_test_confusion_matrix.png
│   ├── shap_global_feature_contribution.png
│   └── shap_local_richmond_upon_thames_2025.png
│
├── ui-prototype/
│   ├── README.md
│   ├── homepage.png
│   └── results.png
│
├── .gitignore
├── README.md
├── model_card.md
└── requirements.txt

To reproduce the analysis, create a Python 3.12 environment, install requirements.txt, open the notebook from the repository root and run all cells in order. No GPU is required.

FundFirst is an educational research prototype and does not provide financial, mortgage, lending or investment advice.