Hi, my name is Preston.

How should a home acquisition operator jointly choose an offer and a resale-price path when seller acceptance, buyer demand, carrying costs, and terminal value are uncertain?

Offer-to-Exit: Pricing a Home Under Uncertainty

Offer-to-Exit is my featured GitHub project and a stylized, finite-horizon model of a home acquisition firm's pricing problem. A higher acquisition offer raises the probability of seller acceptance but compresses the margin conditional on purchase. A higher resale price raises proceeds conditional on sale but lowers the weekly sale hazard and increases expected carrying costs. Because the value of the acquisition depends on the optimal exit policy, the two pricing margins must be solved jointly.

The operator chooses an acquisition offer and, conditional on acceptance, a weekly resale-price path over 17 weeks. The empirical components are a predictive distribution for terminal resale value, a binary-choice model of seller acceptance, and a discrete-time duration model of sale with right censoring. The resale problem is solved by backward induction, and its continuation value enters the acquisition choice. The objective maximizes expected contribution profit net of a configurable 95% CVaR penalty, subject to margin, sell-through, loss-probability, price, and markdown-cadence constraints. Declining to price is the outside option; human review is a separate support rule for observations with invalid or weakly supported inputs.

The identification problem is that acquisition offers and list prices are endogenous: offers respond to seller and property characteristics, while markdowns respond to latent demand. Public records do not reveal acceptance under offers never made or buyer demand under prices never posted. The released experiment therefore uses a Monte Carlo data-generating process in which offer-to-value ratios and list-price premia vary experimentally and the true behavioral response functions are known. Models are fit in one generated environment and evaluated in an independently seeded environment with a documented covariate shift.

The repository demonstrates that I can formulate a dynamic firm problem; distinguish predictive objects, causal-response parameters, and policy functions; enforce the decision-time information set; estimate binary-choice and duration models; and carry uncertainty into a constrained decision rule. A separate public-data pipeline for Phoenix and Maricopa County, automated tests, versioned results, model and data cards, limitations, and a decision explorer make those claims inspectable. The evidence supports recovery of known simulated responses and disciplined policy behavior. It does not estimate real-market elasticities or profit lift.

View the GitHub repository and evidence

The Heterogeneous Impact of a Homicide on Nearby Property Values Across Demographic Characteristics

Violent crime is a neighborhood disamenity, but the raw correlation between crime and house prices does not identify capitalization because crime moves with persistent neighborhood attributes that also determine prices. The paper treats the precise timing and location of a homicide as a plausibly exogenous shock to perceived safety, conditional on local time effects and observed housing and incident characteristics. The economic object is the market's capitalization of newly revealed risk into transaction prices, including how quickly that response decays with distance.

Identification comes from a spatial difference-in-differences design using 73,082 arm's-length single-family sales in Miami-Dade County from 2010 to 2018. The analysis compares the change in log sale prices within 0.1 miles of a homicide with contemporaneous changes in the 0.1-to-0.2 and 0.2-to-0.3 mile rings. The event window spans two years before and two years after the incident. Census-block-by-year fixed effects absorb common neighborhood-year shocks, while detailed property and homicide controls address observable composition. The identifying assumption is local parallel trends: absent the homicide, transaction prices in the inner and outer rings would have evolved similarly. A triple-difference extension estimates how this local response varies with victim and offender characteristics.

The paper tests the credibility of that counterfactual rather than treating it as automatic. In the pre-period, price and housing-characteristic differences across rings are small and statistically insignificant after controls, and graphical trends are broadly parallel. Assigning false event dates one and two years earlier produces null effects. Restricting the sample to sales inside the 0.3-mile event area and replacing block-by-year fixed effects with homicide-area-by-year fixed effects yields a similar estimate. The preferred result is a 4.8% decline within 0.1 miles, with no detectable decline in the next two rings. The estimand is the effect on observed transaction prices, not the value of every home; selection into sale and external validity beyond Miami-Dade remain important limits.

Read the paper (PDF)