Model Building
How do you build machine learning models that reliably predict price?
Model Training
Barker's machine learning models are updated on a daily basis to capture transaction data as it becomes available. This enables Barker's systems to be up to date on recent market trends.
An example of why this is so important can be seen below in this watch index chart from WatchCharts.com. This index chart shows a steep decline in watch valuations around the middle of March 2022.
Barker's machine learning models automatically capture recent market trends and use these as inputs to predictive machine learning models.
Barker's models are trained and evaluated on predicting product valuations ahead of the training data’s most recent transaction date. This allows the model to find features that influence the future price, and use these features to raise or penalize the prediction - such as recent market downtrends, which would reduce the valuation.
This concept can be extended to other features which may be predictive of near future resale market trends, such as inflation, SP 500 performance, or more.
Model Validation & Quality Assurance
Model runs are always trained and selected based on machine learning best practices. These include methods such as time series cross validation on backtests, quality assurance for model drift, logging of all in and out of sample metrics, and storing metadata to enable lookbacks.
Barker's models are built to allow a high level of explainability as each product’s appraisal can be explained by the characteristics of that product. Model explainability is part of Barker's validation process and is essential for enabling data driven decisions.
Last updated
