Daniel Hardesty Lewis has spent a career reasoning about systems that are inherently uncertain — climate models, geomorphology simulations, disaster resiliency scenarios on supercomputers. His new startup, Homecastr, applies that same instinct to a question most real estate platforms treat as having a single correct answer: what is this house worth in five years?
Launched with coverage of more than one million US residential properties — concentrated in Texas, New York, and Florida — Homecastr doesn't give you a number. It gives you three: a P10, P50, and P90 representing the pessimistic, median, and optimistic price trajectories for a given home over a five-year horizon. That framing, borrowed from ensemble forecasting in climate science and probabilistic risk modeling in quantitative finance, is a deliberate break from the point-estimate approach Zillow's Zestimate made mainstream.
[QUOTE NEEDED — Lewis on why he carried probabilistic methods into real estate, and what single-number forecasts miss for institutional buyers]
The target customer is not someone shopping for a first home. Homecastr is pitching SFR acquisition teams scoring buy/hold/sell decisions across portfolios ranging from 50 to 5,000-plus properties, mortgage risk desks stress-testing collateral under adverse rate scenarios, and investment committees that want forward-looking underwriting data rather than backward-looking appraisals. The self-serve API — instant key generation, sub-second JSON responses, no sales call required — signals a developer-first go-to-market common in data infrastructure startups.
Forecast attributions are surfaced alongside each projection, identifying which macro and local drivers are moving a given property's outlook. That interpretability layer is likely the sharper competitive wedge against incumbent AVM providers like CoreLogic, HouseCanary, and First American — which produce numbers, frequently without explaining them — than the probabilistic framing alone.
Lewis's path to residential real estate is genuinely unusual. He led a $40 million disaster resiliency initiative at the Texas Advanced Computing Center, ran climate simulations on TOP500 supercomputers, and contributed to geomorphology research that earned a Bagnold Medal recognition. He is currently cross-enrolled at Columbia in Urban Planning and Engineering (Machine Learning). His 2023 company, Summit Geospatial, produces high-resolution elevation data for Texas — a hint at the spatial infrastructure instincts he has carried into Homecastr.
On accuracy, the company benchmarks forecasts using median absolute percentage error, citing roughly 8% annual compounding error in its best-performing geographies. That figure needs context: established AVM providers typically report median errors in the 3–6% range, but for current valuations — a materially different, and considerably easier, task than a five-year forward projection. Without a published backtest methodology, the MdAPE claim is hard to evaluate independently, and Homecastr has not released one yet.
*Editor's note: Homecastr is a real estate forecasting platform, not an AI agent in the sense this publication typically covers. We are running this piece because ensemble simulation and probabilistic output design are patterns directly relevant to how agent-based systems in adjacent verticals are being architected, and because Lewis's technical background is worth tracking. This is an editorial exception; readers here for agent-platform coverage should note the distinction.*