PropertyIQ

Enterprise-style Property Valuation & ML Drift Monitoring Platform

PropertyIQ is an end-to-end system that combines:

• live property valuation
• model drift detection
• confidence scoring
• prediction explainability
• city-level trend analytics

for risk-sensitive credit workflows.

---

Table of Contents

• 1. Executive Summary
• 2. Business Problem
• 3. Solution Overview
• 4. Key Results
• 5. Product Interfaces
• 6. System Architecture
• 7. Data Design
• 8. ML Pipeline
• 9. Feature Engineering
• 10. Modeling Strategy
• 11. Drift Detection Framework
• 12. Explainability Framework
• 13. Forecasting Method
• 14. API Contract
• 15. Tech Stack
• 16. Repository Structure
• 17. Local Setup
• 18. Reproducibility & Data Integrity
• 19. Limitations
• 20. Roadmap
• 21. Screenshots
• 22. License
• 23. Contact

---

1. Executive Summary

PropertyIQ addresses a practical lending problem: valuation models trained on historical data degrade when market distributions shift.
The platform continuously evaluates model reliability and translates statistical signals into decision-ready outputs for analysts and loan officers.

Outcome: strong valuation performance on validation data and explicit drift alerts on out-of-period data.

---

2. Business Problem

In collateral-backed lending, incorrect valuation affects downstream credit decisions (e.g., LTV assessment).
When market conditions shift, models can remain in production without obvious failure flags.

Risk if unmanaged

• silent valuation bias
• inaccurate risk pricing
• weak early-warning visibility
• delayed retraining decisions

---

3. Solution Overview

PropertyIQ provides two integrated surfaces:

Module	Primary User	Purpose
Admin Console	Risk Analyst	Monitor drift, model health, error trends, alerts
Client Portal	Loan Officer	Generate valuation, view trust score, inspect explanation

Core capabilities

• Continuous drift diagnostics (KS + Chi-square + rolling MAPE)
• Ensemble-based confidence scoring
• Trust-tier translation for non-ML users
• SHAP-style explainability outputs
• City-level forecast context

---

4. Key Results

Metric	Value	Target
Sale Price MAPE	1.61%	< 15%
Rental Value MAPE	19.64%	< 25%
OOB R^2 (Sale)	0.9968	> 0.95
Total Records	600,000	--
Cities	10	--
API Endpoints	14	--
Max City KS	0.5566 (Hyderabad)	> 0.30 drift threshold
All City p-values	< 0.05	drift detected

---

5. Product Interfaces

5.1 Admin Console

• Drift summary and severity
• Feature-level KS ranking
• City-level KS ranking
• Rolling MAPE windows
• Chi-square categorical drift
• Global feature influence view
• Alerts with recommendations

5.2 Client Portal

• Input property details
• Live model valuation
• Confidence score
• Trust tier:
  • TRUSTED
  • CAUTION
  • FIELD VERIFICATION
• Top prediction drivers
• Locality map
• City forecast panel

---

6. System Architecture

                        +--------------------------------------+
                        |          Data Generation             |
                        |  properties_2020 / properties_2025  |
                        +--------------------------------------+
                                         |
                                         v
+----------------------------------------------------------------------------+
|                        Notebook / Offline ML Pipeline                      |
|  NB01 Inspection -> NB02 Preprocess -> NB03 Features -> NB04 Train Models |
|  NB05 Drift Analysis -> NB06 Explainability/Forecast Params                |
+----------------------------------------------------------------------------+
                                         |
                                         v
                     +----------------------------------+
                     |  Artifacts + JSON Outputs       |
                     |  *.pkl, encodings.json,         |
                     |  drift_results.json, shap.json  |
                     +----------------------------------+
                                         |
                                         v
+----------------------------------------------------------------------------+
|                             FastAPI Backend                                |
|  /api/predict/*  /api/drift/*  /api/model/*  /api/forecast/* /api/shap/*  |
+----------------------------------------------------------------------------+
                                         |
                                         v
                    +--------------------------------------+
                    |         React Frontend              |
                    |  Admin Console + Client Portal      |
                    +--------------------------------------+

6.1 Request Flow (valuation)

1. UI sends request to POST /api/predict/valuation
2. API loads cached model + encoding artifacts
3. API builds 14-feature vector in fixed order
4. RF predicts valuation
5. Tree-level variance converted into confidence score
6. Trust tier assigned and response returned to UI

---

7. Data Design

Two synthetic datasets, macro-anchored and reproducible:

File	Year	Rows	Role
properties_2020.csv	2020	300,000	Train + validation source
properties_2025.csv	2025	300,000	Drift evaluation window

7.1 Macro anchors used

• Housing index trend references
• Policy rate trend references
• City-wise growth priors
• Fixed random seeds (2020, 2025) for reproducibility

7.2 Coverage

Mumbai, Delhi, Bengaluru, Hyderabad, Pune, Chennai, Kolkata, Ahmedabad, Gurgaon, Navi Mumbai.

---

8. ML Pipeline

Notebook	Stage	Output
NB01	Data inspection	inspection_report.json
NB02	Preprocessing + splits	train/val/drift partitions
NB03	Feature engineering	encodings.json
NB04	Model training	sale_price_v1.pkl, rental_value_v1.pkl
NB05	Drift detection	ks_results.json, drift_results.json
NB06	Explainability + forecast	shap_values.json, forecast_params.json

---

9. Feature Engineering

Final sale model uses 14 engineered features:

1. bhk
2. total_sqft
3. bath
4. bath_per_bhk
5. sqft_per_bhk
6. is_large_property
7. city_median_price_sqft
8. locality_median_price_sqft
9. price_sqft_city_zscore
10. city_tier_encoded
11. demand_supply_ratio
12. rbi_hpi_avg
13. interest_rate
14. livability_score

Design choices

• Target-style encoding for city/locality price signals
• City-aware engineered context (z-score positioning)
• No scaling needed for tree-based model

---

10. Modeling Strategy

10.1 Sale Model

• RandomForestRegressor
• n_estimators=300
• max_depth=10
• min_samples_leaf=4
• oob_score=True

10.2 Rental Model

• Separate target and feature emphasis
• Includes rental-relevant factors (amenity/furnishing context)

10.3 Why Random Forest

• Handles non-linear interactions
• Works well on mixed tabular features
• Robust with limited heavy preprocessing
• Enables ensemble disagreement-based confidence signal

---

11. Drift Detection Framework

11.1 Statistical methods

• KS-test for continuous features
• Chi-square for categorical features
• Rolling MAPE for temporal degradation tracking

11.2 Severity interpretation (KS)

KS Statistic	Interpretation
0.00 - 0.10	Minimal shift
0.10 - 0.30	Moderate shift
0.30 - 0.60	Significant drift
> 0.60	Extreme drift

11.3 Observed highlights

• Major macro feature drift detected
• Strong city-level heterogeneity
• Tail-window error spikes visible in rolling MAPE

---

12. Explainability Framework

PropertyIQ returns:

• Global feature influence summaries
• Per-request contribution drivers

Observed behavior in runs:

• Location-derived features dominate physical features for valuation impact in this dataset context.

---

13. Forecasting Method

City-level forecasts use implied annual growth from historical medians:

$$ \text{CAGR} = \left(\frac{\text{median}_{2025}}{\text{median}_{2020}}\right)^{1/5} - 1 $$

Projected values are generated across horizons with confidence bands for UI consumption.

---

14. API Contract

14.1 Drift

• GET /api/drift/summary
• GET /api/drift/ks-features
• GET /api/drift/ks-cities
• GET /api/drift/rolling-mape
• GET /api/drift/chi2
• GET /api/drift/city/{city}

14.2 Forecast

• GET /api/forecast/all
• GET /api/forecast/{city}

14.3 Model

• GET /api/model/registry
• GET /api/model/shap
• GET /api/model/localities
• GET /api/model/localities/{city}

14.4 Prediction / Explainability

• POST /api/shap/explain
• POST /api/predict/valuation

14.5 Sample request/response

{
  "city": "Mumbai",
  "locality": "Bandra West",
  "bhk": 3,
  "bath": 3,
  "total_sqft": 1200
}

{
  "predicted_price_sqft": 21067.0,
  "confidence_score": 81.0,
  "trust_tier": "CAUTION",
  "top_drivers": [
    "locality_median_price_sqft",
    "city_median_price_sqft",
    "price_sqft_city_zscore"
  ]
}

---

15. Tech Stack

Layer	Tools
Backend	Python 3.13, FastAPI, Uvicorn, scikit-learn, SciPy, Joblib
Frontend	React 18, Vite, Tailwind CSS, Recharts, Framer Motion, React Leaflet
Data/Graph	Neo4j
Serialization	JSON + Joblib artifacts

---

16. Repository Structure

dashboard/
|- api/
|  |- main.py
|  |- config.py
|  |- models/
|  \- routes/
|- frontend/
|- notebooks/
|- models/
|- outputs/
\- README.md

---

17. Local Setup

Prerequisite: run NB01 -> NB06 to generate required artifacts.

Backend

cd dashboard/api
pip install fastapi uvicorn pandas numpy scikit-learn joblib scipy
python -m uvicorn main:app --reload --port 8000
# Docs: http://localhost:8000/docs

Frontend

cd dashboard/frontend
npm install
npm run dev
# App: http://localhost:5173

---

18. Reproducibility & Data Integrity

• Fixed seeds for data generation and split control
• Train/validation sourced only from 2020 data
• 2025 data isolated as drift window
• Encoding artifacts fit on training data only
• API uses persisted artifacts for deterministic inference

---

19. Limitations

• Synthetic dataset (macro-anchored) rather than proprietary bank transaction logs
• Confidence score is ensemble-variance based, not full predictive interval modeling
• Forecasting uses simplified growth assumptions
• Production security/compliance hardening not included in local demo mode

---

20. Roadmap

• Model registry with version promotion policy
• Scheduled drift jobs + notification pipeline
• Retraining trigger automation with governance thresholds
• Role-based access and audit logging
• Containerized deployment + CI/CD
• Monitoring stack (latency, error, drift, prediction quality)

---

21. Screenshots

21.1 Landing Page

Landing page screenshots (use the image files below):

21.1.1 Hero

21.1.2 Admin Preview Block

21.1.3 Problem Section

21.1.4 Three Layers Section

21.1.5 Metrics Section

21.1.6 Comparison Table

21.1.7 RBI Alignment Section

21.2 Admin Console

Admin Console screenshots (use the image files below):

21.2.1 Drift Monitor Overview

21.2.2 City-Level KS Drift Table

21.2.3 Model Health

21.2.4 City Analytics

21.2.5 Alert Log

21.3 Client Portal

Client Portal screenshots (use the image files below):

21.3.1 Empty State

21.3.2 Valuation Result (Field Verification)

21.3.3 Map and Market Context (Mumbai)

21.3.4 Forecast View (Mumbai)

21.3.5 Valuation Result (Trusted)

21.3.6 Map and Market Context (Pune)

21.3.7 Forecast View (Pune)

---

---

23. Contact

• Name: Pranav Panchal
• LinkedIn: https://www.linkedin.com/in/pranavpanchal1326/
• GitHub: https://github.com/pranavpanchal1326