Across all markets, our forecasting models explain more than 95% of the past variability in price changes (R2>0.95). Our models yield highly accurate out-of-sample forecasts in the near-term and long-term, and they continue to improve through our ongoing research and development.
MdAPE is short for median absolute prediction error, and is a statistical measure of the valuation model accuracy. The valuation MdAPE is computed as follows:
Step 1: In any given month, HouseCanary produces and stores a set of 100m property valuations nationwide.
Step 2: Over the next month we obtain actual prices of arms length transactions occurring nationwide.
Step 3: We measure the absolute percent deviation of actual sales prices to what we had estimated the property values to be prior to the observed sale.
Step 4: Finally, the MdAPE is computed as the median absolute percent deviation of all transactions observed nationwide.
In summary, A lower MdAPE indicates a more accurate valuation.
For every market that we forecast nationwide, we go back 240 months from today. We cut and store the data at each month to simulate the information available at that point in time.
Next, we let the forecast model run on the cut data, and produce a forecast as of that date in time.
Finally, we take the 12 month forecast obtained using the cut data, and measure it against the known index values held out from the cut sample. The MdAPE provided is the median 12 month forecast error observed after letting the system run across every market over the past 20 years.
We've built an advanced analytics platform to understand why changes happen and forecast real estate values. We use these forecasts to help you make informed business decisions.
Accurate forecasts and insights begin with great data. We have built an advanced real estate data platform to predict future real estate prices, including 1,000+ data series for each real estate market and 50+ years of monthly history. Our real estate data is combined with macro-economic data, capital markets data, mortgage data, search & social data, and house/parcel data to form the most comprehensive and multi-faceted data set.
We transform data to develop "smart" proprietary variables that help to explain market behaviors over time and eliminate spurious relationships. Our "smart" proprietary variables explain how the market behaves and where we are in the housing cycle.
For each metro area and market, we identify which variables are driving future change. We use machine learning to define how far each variable leads what we are predicting and the relationship. This enables us to leverage data that is known as of today, that has leading relationships to create time-tested predictive models.
We turn data into predictive models you can trust. Using our leading indicators, we build customized models for each metro area and zip code that best explain price change 36 months into the future. Using advanced statistical methods, machine learning, and mass computing power, we scan the space of up to 4 billion potential models for each area, and average the best ~1,000 models to arrive at the most accurate forecast.
Every month we back-test our models for each metro area and zip code to see how our predictions match up with what actually happened. Across all markets, our models explain more than 95% of the past variability in price changes (R2>0.95). Furthermore our models yield highly accurate out-of-sample forecasts in the near-term and long-term, and they continue to improve through our ongoing research and development.