Housing Market Forecasts: Accurate Forecasts Combining Machine Learning with Expert Models

Elder Research worked with a housing manufacturer to build fast and accurate forecasts of housing-related economic indicators. These forecasts combined machine-learning methods with expert models to produce high-quality predictions that outperformed benchmark methods under rigorous testing. The models were also designed for interpretability, empowering business users to understand what data are used and how to make specific predictions.

Because this client operates in the housing industry, their business is closely tied to the housing market. Their planning and projections depend on accurate housing-activity forecasts reaching years into the future. Historically, these projections have been compiled from several sources including paid market analysis, simple market models, and a strong understanding of the market and the client’s position within it.

Recognizing the potential in bringing machine learning (ML) to bear on this problem, the client tasked Elder Research with building a new collection of market forecasts.

The goal: profitably using the wide range of data available to the client, both public and private, to produce more accurate predictions than have been possible historically.

The client also asked Elder Research to explore applications of natural language processing (NLP) to their collection of document-based data sources. This would programmatically extract relevant information otherwise locked inside these documents and relieve the client’s team of tedious manual efforts.

Elder Research Forecast matches or outperforms competitive forecasts at all prediction horizons.

Elder Research developed a suite of housing-related forecast models, applying our expertise to optimize accuracy, minimize variability, and rigorously test our methodology. To optimize accuracy, we combined tried-and-true subject-matter expertise—simple models based on economically sound understanding—with wide-ranging and flexible ML models capable of synthesizing thousands of inputs to make decisions.

To reduce variability, we combined multiple high-performing models into model ensembles, allowing our forecasts to combine different perspectives for a given task. And, as always, we deployed a

Explainable results show which features had the greatest impact on prediction

rigorous model-testing strategy to measure how our modeling framework would have performed had it been developed in previous years. This work was then collected in a Python package, making it possible to deploy the framework and attach a user interface.

Elder Research also leveraged NLP and programmatic text processing to identify and extract key information from several of the client’s document-based sources. This extraction pipeline made it possible to compare our model predictions—and past performance under our testing strategy—to actual expert performance. This pipeline can also serve as a baseline for future NLP efforts.

Elder Research’s ensemble market models matched or outperformed the forecasts contained in the client’s paid subscriptions over more than a decade of testing. This demonstrated the effectiveness of combining machine learning and expert approaches.

Our forecasting models were also designed with interpretability in mind, providing insight into how their various inputs are used and empowering business users to understand why models make specific future predictions. This transparency allows users, for example, to understand how our models’ views into the market change as they are asked to predict further into the future.