Full-Stack AI Asset Price Forecasting Platform for Leasing
The Problem
A European leasing company was making multi-year financial commitments on vehicle portfolios without reliable long-term price forecasts. Every lease contract depends on a residual value assumption — what the vehicle will be worth when the lease ends in three, five, or seven years. Too high, and the company absorbs the loss on resale; too low, and lease payments are uncompetitive. Their existing approach relied on quarterly industry guidebooks and analyst judgment, which degraded rapidly beyond two years — and with commercial lease terms reaching seven years, the exposure was significant. A single percentage point of error across thousands of vehicles meant millions in unexpected losses or missed revenue. They needed a system that could predict vehicle prices seven years out with quantified accuracy, cover their full portfolio of makes and configurations, and present results in a dashboard their risk and investment teams could use without data science expertise.
Why Building a 7-Year Vehicle Price Forecasting Model Is Hard
Long-horizon asset price prediction combines the complexity of financial modeling with the peculiarities of the vehicle market — where dozens of variables interact non-linearly over timeframes that exceed most ML training windows:
- Seven-year prediction horizons push beyond most training data — most historical price datasets cover three to five years, so the model must extrapolate beyond its densest training region, where small errors in depreciation curves compound
- Non-linear depreciation curves that vary by segment — a luxury sedan, a commercial van, and an electric vehicle each depreciate differently, influenced by brand perception, demand, and technology obsolescence a single parameter set cannot capture
- Market regime changes over long horizons — in seven years, diesel regulations, EV adoption, supply chain disruptions, and economic cycles can reshape the depreciation curve; the model must account for structural shifts, not just historical patterns
- Configuration-level granularity — the same make and model year can have very different residual values by engine type, trim, mileage, and equipment; make-model-level forecasting misses the configuration effects that drive real exposure
- Accuracy requirements measured in basis points — leasing prices are set on residual assumptions, so an MAE improvement from 0.05 to 0.025 is the difference between competitive rates and margin erosion across thousands of contracts
- Interpretability for risk and investment decisions — risk managers cannot use a black box; they need to understand what drives each forecast to make informed portfolio decisions
What We Did
Data Engineering & Feature Pipeline
- Built the historical vehicle price data pipeline in Python — ingesting, cleaning, and normalizing years of European transaction data across makes, models, trims, and configurations
- Developed feature engineering capturing the variables driving long-term depreciation — age, mileage, engine type, transmission, trim, brand segment, MSRP, supply indicators, and macroeconomic factors
- Implemented temporal feature construction and data quality validation — encoding each variable's trajectory over time and flagging implausible or duplicate records with domain-informed imputation
Stacked Regression Model Architecture
- Designed the stacked regression ensemble using scikit-learn — combining gradient boosted trees, random forests, ridge regression, and SVR into a meta-learning architecture that optimally weights each base model
- Implemented segment-specific model training for luxury, economy, commercial, and electric vehicles, capturing segment dynamics while sharing cross-segment features through the stacking layer
- Tuned hyperparameters with cross-validated and Bayesian optimization and validated against MAE, RMSE, MAPE, and business metrics — achieving MAE 0.025 on normalized prices across the full horizon
Streamlit Dashboard & Forecasting Interface
- Built the interactive forecasting dashboard in Streamlit — risk managers select vehicles by make, model, configuration, and lease term and see predicted price trajectories with confidence intervals
- Implemented portfolio-level analysis — users upload their portfolio and the dashboard generates aggregate residual forecasts, highlighting the highest-risk vehicles and total exposure at each future year
- Developed scenario modeling and forecast tracking — adjusting assumptions (fuel prices, EV adoption, growth) for stress testing and comparing predictions to previous forecasts and realized values
Risk Management Tools & Deployment
- Developed risk scoring that translates predictions into risk tiers based on forecast uncertainty, depreciation velocity, and sensitivity to assumptions, prioritizing attention where exposure is greatest
- Built alerting, a retraining pipeline, and export/reporting — flagging trend shifts, periodically retraining the ensemble, and exporting board-ready reports traceable to underlying predictions
- Deployed with role-based access — portfolio managers see their segments, risk managers see aggregate exposure, and executives see summary dashboards with appropriate drill-down
Key Results
In Their Words
Trembit built us a forecasting system that predicts vehicle prices seven years out with accuracy our analysts could not match manually. The Streamlit dashboard means our risk team actually uses it every day — they do not need to ask data scientists to run a query. It has changed how we price leases.
Their proactive team gets things done as if it were their own project.
What We Learned
Stacked regression outperforms any single model for long-horizon asset prediction
Gradient boosted trees capture the non-linear inflection points (the warranty-exit cliff), ridge regression captures the smooth long-term trend, and random forests handle feature interactions. No single model does all three well. The stacking layer learns which base model to trust at which point in the horizon and for which segment, maintaining 0.025 MAE from year one through seven — where individual models degraded significantly beyond year four.
Feature engineering matters more than model complexity for vehicle price prediction
Deep neural networks with raw features achieved respectable but not competitive accuracy. Switching to a stacked ensemble of simpler models with heavy feature engineering — depreciation rate features, brand strength indices, segment demand ratios, regulatory impact scores — improved accuracy by over 40%. Encoding expert knowledge about how vehicles depreciate gave the simpler models the signal they needed.
A forecasting model risk managers cannot interrogate will never be adopted
The first dashboard showed predictions and confidence intervals — statistically complete but operationally useless for explaining to the board why an assumption changed. We added feature importance breakdowns for every prediction, and adoption climbed dramatically. Risk managers stopped treating the model as a black box to override and started using it as a reasoning tool that amplified their judgment.
Need AI Price Forecasting?
Book a 30-minute architecture session — we'll discuss your forecasting requirements and the infrastructure decisions that matter most. No pitch deck. Just engineering clarity.