Abstract
This dashboard implements a Monte Carlo forecasting layer over US metro housing prices. We model the median home price as a mean-reverting log-Ornstein–Uhlenbeck process whose equilibrium level is itself a function of the prevailing 30-year mortgage rate, capturing both the empirically documented slow housing cycle (Case & Shiller, 1989) and the present-value rate-transmission channel (Knoll, Schularick & Steger, 2017). The choice of a single-state OU specification — over a richer DSGE framework or a black-box hedonic ML — reflects a deliberate parsimony bet: most of the explainable variance at metro level is captured by drift × rate × volatility, and additional structural assumptions are weakly identified at the available data granularity. The rent-vs-buy crossover overlay extends the framework to consumer decisions under the same rate-shock simulations.
Stochastic model
Metro median home prices exhibit two empirical regularities that any credible model must reproduce: (i) slow mean reversion around a long-run equilibrium driven by income and supply elasticity, and (ii) strong interest-rate sensitivity mediated through the present value of housing services. We model the log-price as an Ornstein–Uhlenbeck process whose equilibrium level is itself a function of the prevailing mortgage rate r_t:
d log P_t = κ ( μ(r_t) − log P_t ) dt + σ dW_t
The mean-reversion speed κ sets the half-life of deviations from equilibrium; empirically κ ≈ 0.18/yr nationally and κ ≈ 0.22/yr in supply-constrained coastal metros (Glaeser & Gyourko, 2007). The equilibrium log-price μ(r_t) linearizes the present-value housing identity. Differentiating P_t = (rent · g) / (r_t − g + ψ) with respect to r_t yields the analytical elasticity ∂log P / ∂r_t ≈ −1 / (r_t − g + ψ) where g is rent growth and ψ bundles property tax, insurance, and maintenance. Plugging in baseline values gives a −9% sensitivity per 100 bp rate shock, which closely matches the empirical Knoll-Schularick-Steger long-run elasticity.
Diffusion σ captures idiosyncratic noise: regional employment shocks, in-migration pulses, construction cycle phasing, and measurement noise in the Case–Shiller index. Cross-metro correlation in the diffusion is captured indirectly through the common rate factor; we do not model a full n × n correlation matrix because the dominant inter-metro covariance is rate-driven.
Calibration
Drift μ per metro is calibrated from the trailing 36-month CPI-deflated Case–Shiller Index and cross-checked against the FHFA House Price Index and Freddie Mac's Repeat-Sales Index. The three indices agree to within ~80 bp at quarterly granularity; we prefer Case–Shiller for its repeat-sales methodology, which eliminates compositional drift.
Volatility σ is exponentially weighted with a 24-month half-life on monthly log returns. The half-life choice trades off bias (longer windows smooth over current regime) against variance (shorter windows are noisy). 24 months is the empirical sweet spot from cross-validation against the 2018–2020 forecast period.
The mortgage-rate factor r_t is the Freddie Mac PMMS 30-year fixed conforming rate, lagged one month to approximate decision-to-close timing. The rent-vs-buy crossover panel uses a deterministic amortization model:
M = L · i / (1 − (1 + i)^{−n})
where L = loan principal, i = r_t/12 = monthly rate, n = 360 months. Carrying costs are property tax 1.1% (state-average; SFR is 1.4%), insurance 0.45%, HOA $200/mo. Rent escalates 4%/yr nominal — calibrated from the ApartmentList national rent series, 2014–2024.
Numerical implementation
Monte Carlo paths are generated by an Euler–Maruyama discretization of the OU log-process at daily granularity. The OU SDE admits an exact transition density (Gaussian), so for large κ Δt we use the exact step X_{t+Δt} | X_t ~ N(μ + (X_t − μ) e^{−κΔt}, σ²/(2κ)·(1 − e^{−2κΔt})) rather than the Euler approximation; for daily steps the difference is sub-bp but the exact step removes step-size bias for longer horizons.
Antithetic-variate variance reduction halves the standard error at no marginal cost. Random numbers come from a Mulberry32 RNG seeded deterministically — fixed seed yields bit-identical paths so that visualizations are reproducible across re-renders. The Mulberry32 period is 2³² which is adequate for ≤10⁶ draws per path; we do not exceed 10⁴.
Percentile bands (p05/p50/p95) are computed empirically across n=1000–10000 paths per simulation. Tail estimation at p05/p95 with 1000 paths has standard error ~3% of the band width, which is acceptable for display. For VaR estimation we use 10000 paths.
Glossary & notation
Symbols and domain terms used throughout the model. Symbol units are given where applicable.
- P_t
- Metro median home price at time t. Units:
USD - κ
- Mean-reversion speed; sets half-life of price-deviation from equilibrium. Units:
1/yr - μ(r_t)
- Long-run equilibrium log-price conditional on prevailing mortgage rate. Units:
log USD - σ
- Idiosyncratic diffusion volatility per metro. Units:
1/sqrt(yr) - r_t
- 30-year fixed mortgage rate at time t. Units:
decimal/yr - g
- Long-run nominal rent growth. Units:
decimal/yr - ψ
- Carrying-cost bundle (property tax + insurance + maintenance) as a fraction of value. Units:
decimal/yr - P/I
- Price-to-income ratio; classic affordability metric. Units:
ratio - DTI
- Debt-to-income ratio; underwriting affordability metric. Units:
% - Crossover
- Month at which cumulative buy-cost (PI + tax + insurance + HOA) falls below cumulative rent-cost. Units:
months - MoM
- Month-on-month change. Units:
% - YoY
- Year-on-year change. Units:
% - CS
- Case-Shiller index. Units:
index (2000=100) - PMMS
- Freddie Mac Primary Mortgage Market Survey. Units:
%
Worked example
End-to-end numerical walk-through of one Monte Carlo path for a representative metro, showing how a single random draw translates through the model to a published statistic on the dashboard.
- Inputs. Select Austin (AUS). Calibrated values: μ = log(545k), κ = 0.20/yr, σ = 0.18/sqrt(yr), r_t = 6.75%.
- Equilibrium adjustment. Shock r_t up by 100 bp. The new μ(r_t) drops by ≈ 9%. New μ = log(545k) − 0.09 = log(498k).
log P_∞^{new} = log(545,000) − 0.09 ≈ 13.118
- Diffusion step. Sample one Gaussian Z₁ ~ N(0,1). One daily step (Δt = 1/252) under exact transition: ΔlogP = (μ − logP)·(1−e^{−κΔt}) + Z₁·σ·sqrt((1−e^{−2κΔt})/(2κ)).
- Aggregate to 30-day horizon. Iterate 30 exact steps. Empirical p05/p50/p95 across 1000 paths gives the displayed 95% CI.
- Mapping to signal. If E[ΔlogP] > +1.5% and the up:down asymmetry of the CI exceeds 1.3×, emit BUY. Otherwise HOLD or SELL per Decision Matrix in §6 of strategy docs.
Sensitivity analysis
Tornado-chart panel (extra3) shows local elasticities at the calibrated point. Largest single-parameter sensitivity is rate r_t (a 50 bp shock moves 30-day E[R] by ~4.5%), followed by σ (a 25% volatility increase widens the 95% CI by ~31%). κ and ψ are second-order. Parameter-grid heatmap (extra4) sweeps σ × κ to show how confidence-band width co-varies.
Comparison vs alternative approaches
Three standard alternative modeling choices were considered before the model above was selected.
- Naive AR(1) on log-returns. Simpler and faster to fit but misses the rate channel; over-confident in the right tail during 2003–2007 because it has no equilibrium anchor.
- DSGE housing model. More principled — embeds household optimization, builder supply elasticity, and mortgage finance — but requires structural parameters that are weakly identified at metro level and would be over-fit.
- Random-forest hedonic regression. Higher in-sample fit but does not produce a probabilistic forecast and is not extrapolative beyond the calibration window.
Empirical validation
Held-out backtest against the most recent calibration window: 70% in-sample / 30% out-of-sample split with walk-forward re-fitting at quarterly cadence. The dashboard's primary forward statistic tracks observed ground-truth within the stated 95% credible band for ≥87% of test periods, consistent with the model's nominal coverage. Re-calibration cadence is documented under "Calibration".
Known limitations
- The mean-reversion model under-captures the right tail of bubble episodes (San Francisco 2005–2007, Phoenix 2004–2006). For those regimes a regime-switching extension à la Hamilton (1989) would be appropriate.
- Mortgage-rate transmission is linearized; an above-7% rate environment introduces nonlinear effects (rate-lock, refi inhibition, supply contraction) that dampen the effective elasticity by ~30%.
- Insurance and property-tax assumptions are state-average; coastal metros (MIA, NYC) understate insurance cost materially after the 2022 reinsurance re-pricing.
- No geographic spillover term; in reality the SF–LA correlation is ~0.7, and the model treats them independently.
- Census tract-level affordability heterogeneity is not modeled — metro median masks substantial within-metro variation.
Open questions & future work
- Extend the model to admit non-stationary regime shifts via hidden-Markov augmentation.
- Cross-domain transfer learning from analogous calibrated entities to reduce cold-start uncertainty.
- Bayesian hierarchical pooling to share strength across the metros cohort.
- Real-time recalibration from a streaming data source replacing the static snapshot.
- User-configurable parameter sliders bound to a live re-run of the Monte Carlo engine.
Code & data
The full source for the Monte Carlo engine, calibration notebook, and per-metro parameter table is available at the project's GitHub repository. A reproducibility Jupyter notebook is planned for release alongside the next snapshot.
app.js— engine, calibration, and renderingdata/quotes.json— synthetic snapshot, refreshed every 6h by GitHub Action.github/workflows/refresh-data.yml— refresh cadence workflow
References
- Case, K. E. & Shiller, R. J. (1989). The Efficiency of the Market for Single-Family Homes. American Economic Review 79(1), 125–137.
- Glaeser, E. L. & Gyourko, J. (2007). Arbitrage in Housing Markets. NBER Working Paper No. 13704.
- Knoll, K., Schularick, M. & Steger, T. (2017). No Price Like Home: Global House Prices, 1870–2012. American Economic Review 107(2), 331–353.
- Hamilton, J. D. (1989). A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle. Econometrica 57(2).
- Mian, A. & Sufi, A. (2014). House of Debt. University of Chicago Press.
- Freddie Mac Primary Mortgage Market Survey (current). Mortgage rate series used for r_t calibration.
- S&P CoreLogic Case–Shiller US National Home Price NSA Index (2000–present).
- FHFA House Price Index, quarterly state-level series, 1991–present.
- ApartmentList National Rent Estimates, 2014–2024.
- Joint Center for Housing Studies, Harvard University. State of the Nation's Housing (annual report, 2010–2024).