Reports
Markets
Latest
Maps
Learn
Chat
This study forecasts owner-occupied and renter household formation at the county level across Indiana over a five-year horizon (2025 to 2029), then translates those forecasts into estimates of housing supply need. The core modeling approach uses a rolling five-year window regression, trained on historical relationships between local economic and demographic conditions and subsequent household growth. The model is evaluated using leave-one-window-out cross-validation, where each test window is predicted using only data from earlier periods.
The following data series are used as model inputs and for computing supply need components.
Separate models are estimated for owner-occupied and renter household formation. Both use the same rolling window approach: for each historical five-year window, the model is fit on anchor-year conditions and asked to predict the change in households five years later. This structure mirrors the forecast task directly.
Owner model
The owner household model regresses five-year change in owner-occupied households on the following anchor-year predictors: mean home sale price; the interaction of sales volume with mean price; one-year change in owner households; one-year change in employment; five-year domestic migration level; five-year international migration level; domestic and international migration rates (flows as a share of total households); and interactions between current owner household count and each migration level term. The model specification in R is:
total_change ~ mean_price + total_sales:mean_price + owner_diff1 + employment_diff1 + dom_mig_5yr + intl_mig_5yr + dom_mig_rate + intl_mig_rate + owner_householdsE:dom_mig_5yr + owner_householdsE:intl_mig_5yrRenter model
The renter model uses the same rolling window structure with a parallel set of renter-specific predictors, including renter vacancy rates, median rent interacted with sales volume, and renter-specific migration and trend terms.
Training and forecasting procedure
For validation, the model is trained on all five-year windows where the anchor year is earlier than the test window. No future data is ever used to train a model that predicts an earlier period. For the production forecast, all valid historical windows are used for training, and the model is applied to 2024 anchor-year conditions to project 2025 to 2029 change. Predicted change is applied purely from the regression; no trend blending or post-hoc adjustment is applied.
Driver contributions reported in the dashboard are computed as the term-level product of each predictor value and its estimated coefficient, then grouped into reader-friendly categories (Market conditions, Migration, Employment, Recent trend, Market size).
Validation uses a leave-one-window-out procedure constrained so that the training period always precedes the test period. Five test windows are evaluated for the owner model (holdout anchor years 2015 through 2019) and four for the renter model (2016 through 2019), each covering all 92 Indiana counties, yielding 460 and 368 county-year observations respectively.
Error categories are defined as follows. A prediction is classified as "Nailed it" if the absolute error is 100 or fewer households, or the percentage error is 10 percent or less. "Pretty close" covers predictions within 500 households or 25 percent of actual. "Right direction, too far" and "Right direction, not far enough" describe predictions with the correct sign but incorrect magnitude. "Wrong direction" cases are those where the model predicts growth when decline occurred, or vice versa.
| Metric | Owner model (460 county-year obs., 5 windows) |
Renter model (368 county-year obs., 4 windows) |
|---|---|---|
| MAE | 510 households | 357 households |
| Median absolute error | 316 households | 221 households |
| RMSE | 889 households | 543 households |
| Correlation (predicted vs. actual) | 0.963 | 0.850 |
| R² | 0.927 | 0.722 |
Error category breakdown — Owner model:
Error category breakdown — Renter model:
The owner model performs well, with nearly 60 percent of county-year predictions landing within 500 households or 25 percent of actual, and a correlation of 0.963 between predicted and observed change. Wrong-direction errors are rare (22 of 460, under 5 percent). The renter model shows lower but still meaningful predictive power, reflecting the greater volatility of the rental market and the sensitivity of renter household counts to short-term economic and migration shocks.
For each county and tenure type, the five-year supply need is calculated as:
Supply needed = Projected new households + Anticipated housing loss − Excess vacancy above equilibrium
Anticipated housing loss is estimated from observed historical loss rates applied to the current stock. Housing loss is estimated using the Census Bureau's methodology for housing unit loss, which applies rates derived from the Components of Inventory Change (CINCH) supplement to the American Housing Survey, differentiated by structure type, age of unit, and Census region. Older homes and mobile homes carry substantially higher loss rates than newer single-family construction. Midwest regional rates are applied throughout.²
Excess vacancy is the difference between actual vacant-for-sale or vacant-for-rent units and the equilibrium count, defined as 1 percent of the owner market and 5 percent of the renter market. Where actual vacancy falls below equilibrium, this term is negative, adding to supply need. Where vacancy exceeds equilibrium, it reduces the net need.
Supply need is reported separately for owner-occupied and rental units. The analyses in this report focus on owner-occupied need, with rental need noted but not examined in detail.
Contact. Questions about methodology or data should be directed to the Indiana Association of REALTORS® ([email protected]).