From 10 Years of Transaction Data to a Binding Real-Time Price Prediction

An AI model learns the relationships between object characteristics and actually achieved prices from historical transaction data. These include properties such as age, condition, market segment, and historical price development. The model detects patterns hidden in the mass of data and derives a price prediction for new, not yet valued objects.

The quality of the prediction depends on the data foundation: the more historical transactions are available and the more consistent the data is, the more precise the prediction becomes. In this project, 10 years of transaction data with thousands of orders per day were available. This breadth and depth enabled a model precise enough for binding price commitments.

A confidence interval indicates the range within which the actual price falls with a defined probability. Instead of a single value, the model delivers a corridor — for example: "The price falls between 3,200 and 4,100 euros with 90 percent probability."

The advantage over a point estimate: uncertainty becomes visible and manageable. A company can use the confidence interval for risk management, for example by issuing a guarantee for a value at the edge of the interval. This makes the financial risk calculable, and a statistical prediction becomes a binding business instrument.

In most applications, an AI prediction remains an indicative value: useful for decisions but without binding effect. The step to commitment requires a model that is precise enough, and a business model that factors in the residual risk. In individual cases, the deviation between prediction and actual price can lead to a bad deal. At scale, however, such approaches are economically sound because overestimates and underestimates balance out statistically. The combination of a precise model and calculated risk turns predictive analytics into a binding pricing model.

LightGBM is a gradient boosting framework based on decision trees. It builds trees sequentially, with each new tree correcting the errors of the previous one. The method is particularly suitable for structured, tabular data with many features.

For price predictions on transaction data, LightGBM is well suited for several reasons: it processes large datasets efficiently, captures complex nonlinear relationships between features, and delivers interpretable results. Compared to neural networks, LightGBM requires less training data and computational resources and is less prone to overfitting on structured data.

The minimum requirement is historical transaction data with three components: object characteristics (what was valued?), offer prices or estimates (what was expected?), and realized prices (what was actually paid?). Depending on the industry, market data, time series, and external data sources are added.

The real preliminary work lies in data cleansing and feature engineering: removing outliers, resolving inconsistencies, deriving meaningful features from raw data. In most companies, the necessary data already exists. What's missing is not the volume of data but its systematic preparation and use.

The quality of a prediction model is evaluated through statistical metrics that capture different aspects of prediction accuracy. The most important metrics in the context of predictive analytics are R², MAPE, and MAE.

R² (coefficient of determination) indicates what proportion of price variation is explained by the model. An R² of over 90 percent means: the model captures over 90 percent of the systematic price differences. MAPE (Mean Absolute Percentage Error) measures the average percentage deviation between prediction and actual price. MAE (Mean Absolute Error) gives the average absolute deviation in euros. The combination of these metrics provides a differentiated picture: R² measures explanatory power, MAPE measures relative accuracy, and MAE measures the economic relevance of the deviation.

The underlying principle is applicable across industries. Wherever historical transaction data exists and prices are determined based on object characteristics, a comparable model can be built: used machinery, real estate, IT assets, B2B surplus stock, commodities.

The prerequisites are the same in every industry: a sufficient database of historical transactions, structured object characteristics, and a defined valuation process. What changes are the features, the data sources, and the industry-specific particularities. The statistical method and the architecture remain identical. Mid-sized companies that have several years of transaction data already have the foundation for AI-powered price determination.

Integration is achieved by containerizing the trained model as a Docker image. The container runs on the customer's existing IT infrastructure and is accessed via an API interface. Existing systems retrieve the prediction as a service without having to modify their own architecture.

In this project, the AI model was integrated into the CarTV systems. Insurers and assessors use the prediction through the same interface they already know. The advantage of the container architecture: the model can be updated, scaled, and monitored independently of the rest of the application. No cloud dependency in production.