RESI
SN46AI-powered real estate data and intelligence, the oracle for property markets
A winner-takes-all competition to predict US home prices. Miners build machine learning models, submit them 31 hours before evaluation, and get scored against sales that happened after submission. No peeking. The best predictor takes 99% of emissions.
// The real estate oracle, built by competition.
RESI is a subnet that incentivizes accurate real estate price prediction. Miners train machine learning models to estimate what US homes will sell for, and validators test those predictions against real sales data that the models couldn't have seen at submission time. The most accurate model wins almost all the rewards.
The simple version: Imagine a daily contest where you guess the price of every house that will sell tomorrow. You lock in your guesses today. Tomorrow, the actual sale prices come in, and whoever was closest wins the pot. RESI runs this contest continuously with ML models instead of human guesses.
Centralized equivalent: Think Zillow's Zestimate or Redfin's price estimates, but built through open competition rather than a single company's proprietary model. Anyone can submit a model and try to beat the current best.
How it works:
- Miners train ONNX models for residential price prediction, upload them to Hugging Face, and commit hashes on-chain. Models must be locked in approximately 31 hours before evaluation.
- Validators download committed models, run sandboxed inference against the last 24 hours of real sales data, and score using Mean Absolute Percentage Error (MAPE). The best model takes 99% of emissions, with copycats explicitly detected and zeroed out.
- The problem it solves: Accurate home pricing is worth billions. Zillow lost $881 million on their iBuying program partly due to bad price predictions. Better models mean better decisions for buyers, sellers, lenders, and investors.
- The opportunity: The US residential real estate market is worth over $45 trillion. Pricing accuracy directly impacts mortgage underwriting, property insurance, investment analysis, and automated home buying.
- The Bittensor advantage: Winner-takes-all with anti-copying creates a genuine innovation race. You can't win by cloning the leader; you have to actually beat them. The 31-hour commit window prevents data leakage and ensures models generalize.
- Traction signals: 162 active miners competing daily. 2,417 holders. 52 commits across 3 contributors. Recent work on spatial robustness checks and anti-burn mechanisms. Dashboard live at dashboard.resilabs.ai with W&B tracking.
Category: Real Estate AI / Prediction Markets | Centralized Competitor: Zillow Zestimate, Redfin Estimate, HouseCanary, CoreLogic
RESI has one of the cleanest incentive designs in Bittensor. The 31-hour commit window, winner-takes-all scoring, and copycat detection create a genuine meritocracy: the only way to earn is to build a better model. No gaming, no sybil attacks, no copying the top miner.
Mechanism:
Every day at 18:00 UTC, validators evaluate all committed models against the previous 24 hours of actual home sales. Scoring is straightforward: 1 minus MAPE. A model predicting within 8.5% of actual sale prices scores 0.915. The highest scorer takes 99% of emissions, with a threshold mechanism that protects pioneers from incremental copycats.
The threshold system is clever: all models within a configurable distance of the best score form a "winner set," but within that set, the earliest on-chain commit wins. This means copying the current leader doesn't help. You have to demonstrably improve to displace them. Detected duplicates receive zero emissions.
Models are exported as ONNX (Open Neural Network Exchange), making them portable and vendor-neutral. Validators run inference in sandboxed environments, ensuring reproducibility. The architecture has three layers: miners train and upload, validators fetch and score, and the Pylon layer handles blockchain interaction.
The codebase is lean: 52 commits across 3 contributors in a 787KB repository. Development is steady at 2-3 commits per week, with recent work from konrad0960 on spatial robustness checks and burn mechanism adjustments. The team also pinned bittensor-wallet==4.0.1 to avoid the compromised 4.0.2 package, showing security awareness.
Market metrics show a small but committed community. At 25,145 TAO market cap with 2,417 holders, RESI has the lowest Gini coefficient we've covered: 0.530. That's exceptionally distributed ownership. HHI of 0.025 confirms minimal concentration. Root proportion of 0.172 means demand is organic.
The 7-day price is down 5.5% with a small net outflow of -205 TAO, but the 30-day picture is positive at +17%. The 90-day return of -11% suggests the subnet went through a correction and is stabilizing. Unrealized PnL of 23,183 TAO means most holders are sitting in profit.
The roadmap targets a national property database, seller intent prediction, a public API ecosystem, and expansion into commercial and international real estate. If RESI's models prove competitive with Zillow's, the data becomes the real asset: a continuously updated, competitively validated pricing oracle for the entire US housing market.
- US-only scope: Currently limited to US residential properties. This constrains the addressable market and makes the subnet vulnerable to US housing market cycles.
- Small development team: 3 contributors with steady but low-volume development (2-3 commits/week). The codebase is lean, but scaling data coverage requires more engineering.
- Data source dependency: Miners scrape from public records, Redfin, Zillow, etc. Changes to these sources' terms of service or anti-scraping measures could disrupt the data pipeline.
- Winner-takes-all concentration: 99% to the winner means most miners earn almost nothing. This could discourage participation if the leading model becomes hard to beat.
