UT Dallas UTDesign Capstone — Spring 2026 — Advisor: Muhammad Ikram
How We Predict
Zillow + FRED
70k+ data points
15 Market Features
Price lags, tiers, mortgage rates
XGBoost Model
Trained on 255 zipcodes
Forecast
±1.4% typical error
Factors driving the 1 month forecast. Tap any factor to learn more.
The median home value from last month. This is the single strongest predictor — where prices were most recently is the best starting point for where they're headed.
The median home value from 3 months ago. Comparing this to the current price reveals short-term momentum — whether prices are accelerating, decelerating, or holding steady.
The median home value from 6 months ago. This captures medium-term trends and helps the model distinguish a temporary dip from a sustained shift.
The median home value from 12 months ago. This provides a year-over-year reference point that filters out seasonal fluctuations.
Median price of homes in the top 35th percentile for this area. When premium homes gain or lose value, the rest of the market often follows.
255
Zipcodes
70k+
Data Points
2001–now
History
Predictions powered by XGBoost, trained on Zillow Home Value Index and Freddie Mac mortgage rate data through March 2026.
Model Accuracy
Our 1-month predictions are typically within 1.4% of actual prices (avg. error: $7,415).
1-Month
1.4%
Excellent
$7,415 avg error
3-Month
2.2%
Good
$11,032 avg error
6-Month
2.6%
Good
$12,609 avg error
Under the Hood: XGBoost
XGBoost (Extreme Gradient Boosting) is a machine learning algorithm that builds hundreds of small decision trees, each one learning from the mistakes of the previous ones. Think of it like getting a second opinion, then a third, then a hundredth — each expert focuses on the cases the previous ones got wrong.
It's one of the most widely used algorithms in industry for structured data problems (like predicting home prices from market stats), consistently winning machine learning competitions and powering real-world forecasting systems.
How gradient boosting works:
Training Data
Historical prices
Tree 1
Initial guess
Residuals
What Tree 1 got wrong
Tree 2
Corrects Tree 1
...
Repeat
Final Prediction
Sum of all trees
Ensemble Learning
Combining many weak models into one strong predictor
Gradient Descent
Each tree minimizes the remaining prediction error
Regularization
Constraints that prevent the model from memorizing noise
Our XGBoost models were tuned with specific hyperparameters that control how the algorithm learns. Here's the configuration and what each setting does:
n_estimators1000Maximum number of trees the model can build
learning_rate0.05How much each tree contributes — lower means more cautious learning
max_depth6How many levels deep each decision tree can grow
subsample0.8Fraction of data rows sampled for each tree (reduces overfitting)
colsample_bytree0.8Fraction of features sampled for each tree
min_child_weight5Minimum data required in a leaf node — higher prevents overfitting
reg_alpha0.1L1 regularization — encourages simpler trees
reg_lambda1L2 regularization — penalizes large predictions
early_stopping_rounds50Stop training if accuracy hasn't improved in this many rounds
Trees actually used (after early stopping)
Although the maximum is 1,000 trees, early stopping halted training once accuracy plateaued:
215
1-Month
213
3-Month
511
6-Month
Every prediction uses exactly 15 data points — 14 from Zillow home value data and 1 from the Freddie Mac mortgage rate survey. These features are grouped into five categories:
Price History
(4 features)Historical price lags that anchor predictions to recent and long-term trends
Market Segments
(6 features)Prices across bedroom tiers and market percentiles to capture cross-segment dynamics
Momentum
(1 feature)Year-over-year price change that captures the overall market direction
Context
(3 features)Temporal and geographic signals like seasonality, long-term trends, and market size
Economic
(1 feature)Macroeconomic data — the 30-year fixed mortgage rate from the Freddie Mac Primary Mortgage Market Survey (via FRED)
Two features (1-month and 3-month price changes) were removed during development because they encoded the target variable — a form of data leakage. The remaining 15 features are verified clean predictors. The mortgage rate was added after confirming it improved accuracy on all three forecast horizons.
When you enter a zipcode, here's what happens behind the scenes:
Collect the latest Zillow Home Value Index (ZHVI) data for the selected zipcode
Extract 14 market features: price lags, bedroom tiers, momentum, and seasonal context
Feed features into the trained XGBoost model to predict future overall ZHVI
Scale the ZHVI prediction to the selected bedroom tier using current price ratios
Apply NAHB-derived bathroom multiplier to adjust for the selected bathroom count
Derive direction signal, confidence score, and expected price range from the forecast
Bedroom scaling formula
The XGBoost model predicts overall ZHVI. To get a bedroom-specific price, we apply the current ratio:
predicted_price = predicted_zhvi × (current_bedroom_price ÷ current_zhvi)We don't use one model for everything. We trained three separate XGBoost models, each specialized for a different time horizon:
1-Month
Short-term forecast. Most accurate — prices rarely shift dramatically in 30 days.
3-Month
Medium-term outlook. Captures quarterly trends and seasonal shifts.
6-Month
Longer-range view. More uncertainty, but useful for planning ahead.
Each model was trained on the same 15 features but optimized for its specific target — the actual home value 1, 3, or 6 months into the future. Adding the mortgage rate improved all three horizons, with the largest gain on the 3-month forecast.
The model was trained on 66,521 training rows from 255 Dallas-area zipcodes, spanning 2001 to March 2026.
66,521
Training rows
Before 2025-01
3,825
Test rows
After 2025-01
Key training decisions:
- Time-based split— all data before Jan 2025 is training data; everything after is test data. We never let the model "peek" at future prices.
- Leakage investigation — we removed 2 features (1-month and 3-month price changes) that were essentially encoding the answer. The remaining 15 features were verified to be clean predictors.
- Data sources — Zillow Home Value Index (ZHVI) for price history, smoothed and seasonally adjusted by Zillow; plus the Freddie Mac Primary Mortgage Market Survey (weekly, via FRED) resampled to monthly averages.
- Early stopping — training halts automatically when 50 consecutive rounds show no improvement, preventing overfitting.
We measure accuracy three ways. Here's what each metric means in plain terms:
MAPE (Mean Absolute Percentage Error): 1.4%
On average, our 1-month prediction is within 1.4% of the actual price. For a $400,000 home, that's roughly $5,480 off.
MAE (Mean Absolute Error): $7,414.89
The average dollar amount the prediction is off by. Unlike MAPE, this doesn't scale with price — it's the raw error in dollars.
RMSE (Root Mean Square Error): $15,874.83
Similar to MAE, but penalizes large errors more heavily. If RMSE is much larger than MAE, it means the model occasionally makes bigger misses — ours are close, which means predictions are consistently reliable.
The "up" or "down" direction signal comes directly from the XGBoost 1-month prediction. If the predicted price is higher than the current price, the direction is "up"; otherwise it's "down."
Confidence is derived from the size of the predicted change — a larger predicted move means more confidence. A near-zero change gives ~50% confidence (essentially a coin flip), while a >5% predicted change approaches maximum confidence.
Confidence scale: