Project Case
Shanghai Housing Price Forecast
An explainable forecasting model for Shanghai monthly housing-price growth.
Role
Data collection / Feature engineering / XGBoost modeling / SHAP interpretation
Stage
Master thesis research, model iteration, and market interpretation
Key Outcome
Built an exploratory XGBoost model and used SHAP to examine momentum, liquidity, policy, and sentiment signals; the reported evaluation has data-construction and validation limits.
Strategy Snapshot
BUSINESS QUESTION
How can housing prices and macro variables be analyzed together?
The case frames a research question around model comparison and feature interpretation. Data interpolation and validation design mean the results do not establish advance prediction of observed market changes.
REPORT EVIDENCE
Model interpretation with validation limits
The thesis reports RMSE 320.06 for the final XGBoost version. SHAP highlights price momentum, M2 liquidity, stock-market signals, and consumer confidence; interpolation and validation design limit interpretation.
STRATEGY SIGNAL
Treat it as exploratory research
Further validation with an information-safe time split and observed monthly data would be needed before investment or policy use.
Background
This thesis reframes Shanghai housing-price analysis as a monthly growth forecasting problem. Instead of only describing long-term market trends, it builds a model that can capture short-term movements using housing prices, macro-financial indicators, policy signals, stock-market variables, and consumer sentiment.
Problem
A black-box forecast is not very useful for market judgment. The project needed to answer two questions at the same time: whether monthly housing-price movements could be predicted more accurately than a traditional linear model, and which variables were actually driving the model's prediction.
Approach
I collected and aligned multi-source data from 2000 to 2024, transformed annual and cumulative indicators into monthly features, built lagged variables for policy, LPR, M2, stock-market and sentiment signals, trained iterative XGBoost models, compared them with linear regression, and used SHAP to interpret feature contribution.
Analysis Process
Reframe the target
The target was changed from raw price level to next-month growth, making it consistent with percentage-based macro and financial inputs.
Build market features
The model combined recent price momentum, M2, LPR, CPI, policy scores, stock-market indicators, consumer confidence, and seasonal variables.
Compare model versions
XGBoost was iterated against a baseline model and evaluated on reconstructed price levels using RMSE and MAE.
Explain the result
SHAP analysis translated model output into market logic: momentum, liquidity, seasonality, and investor sentiment became interpretable drivers.
Key Evidence
FINAL MODEL
RMSE 320.06 / MAE 176.82
The final XGBoost version tracked Shanghai monthly housing prices from 2021 to 2024 with much lower error than the baseline.
EVALUATION CONTEXT
RMSE 320.06
The report compares iterations after changing the target and reconstructing price levels. The headline percentage is omitted because the baseline conditions are not sufficiently comparable for a general improvement claim.
TOP SIGNAL
Price momentum
SHAP results showed that previous-month growth was the most influential predictor, indicating market inertia in short-term price movement.
MACRO DRIVER
M2 liquidity
Money supply emerged as an important macro-financial predictor, linking liquidity conditions to housing-market expectations.
Reported Model Metrics
Metrics are reported on reconstructed price levels. Monthly interpolation and validation design limit interpretation; treat these as thesis results, not deployable forecasting performance.
Final XGBoost
RMSE 320.06 · MAE 176.82
Reported metrics on reconstructed price levels; interpolation and validation design limit interpretation.
Initial model
RMSE 18,222.69
Different target and feature setup; not a like-for-like benchmark.
Same-input OLS
RMSE 761.82 · MAE 533.38
Reported comparison; further information-safe time-split validation is required.
Market Driver Map
The model was presented as a market-explanation framework, not only a prediction engine.
Momentum
Previous-month growth
The strongest short-term predictor in SHAP analysis.
Liquidity
M2 money supply
A macro-financial signal linked to market expectation.
Sentiment
Consumer confidence / stock market
Signals investor mood and capital reallocation pressure.
Decision Logic
Validate the information set and out-of-sample performance first
The thesis does not establish a usable market-warning tool. Further work should exclude information unavailable at the forecast date and evaluate on observed monthly values.
Treat SHAP as a model explanation, not a causal result
Feature importance describes how the model uses its inputs; it does not show that those variables caused housing-market changes.
Findings
The thesis produced an exploratory XGBoost model and SHAP analysis. Reported metrics and feature patterns are research findings; interpolation and validation design do not establish production readiness or reliable live-market forecasting.
Key Outputs
- Monthly feature matrix covering housing, macroeconomic, policy, financial-market, and sentiment variables
- XGBoost forecasting model for next-month housing-price growth
- Benchmark comparison with linear regression
- RMSE / MAE evaluation on reconstructed price levels
- SHAP feature-importance interpretation
- Market explanation of momentum, liquidity, and sentiment effects
Reflection
For job-facing presentation, this is my strongest data case: it shows that I can move from messy real-world data to model design, performance evaluation, interpretability, and business-facing market judgment.
Original Deliverable
This page is a concise case summary. The full report contains the research process and modeling details.
View Full Report