Meridian Labs · AI Pricing Engine

Dynamic Pricing
Used Car Valuation Model

Client: Rusty Bargain · 354K vehicles · 3 models benchmarked · LightGBM recommended
Dataset 354,369
Features 8 used
Split 80/10/10
Winner LightGBM
Best REQM
€1,670
LightGBM · test set
▼ best in benchmark
Prediction Latency
0.25s
LightGBM · 29,516 records
near real-time
Training Speed
12.9s
LightGBM · 236K records
393x faster than RF
Dataset Volume
354K+
vehicles after cleaning
262 duplicates removed
REQM Improvement
−37%
LightGBM vs Linear Reg.
1670 vs 2637
🏆
Recommended Model — LightGBM
After benchmarking 3 models across 354K vehicle records, LightGBM is the clear recommendation for Rusty Bargain's pricing app. It delivers the lowest REQM (€1,670), trains in 12.9s — 393x faster than Random Forest — and predicts in under 0.25s, enabling a real-time user experience in the application.
LightGBM WINNER
€1,670
REQM
1,670.50
Train Time
12.9s
Pred. Time
0.25s
Best Params
n=200 / l=50
Random Forest 2ND
€1,727
REQM
1,727.38
Train Time
85s
Pred. Time
0.251s
Best Params
n=100 / d=20
Linear Regression BASELINE
€2,637
REQM
2,636.53
Train Time
0.92s
Pred. Time
0.08s
Type
OHE encoded
REQM Comparison
Lower = better prediction quality · test set results
QUALITY
−3.3%
LightGBM vs Random Forest
1,670 vs 1,727 REQM
−37%
LightGBM vs Linear Reg.
biggest accuracy leap
Training Time Comparison
Seconds to train on 236K records
SPEED
393x
LightGBM faster than RF
12.9s vs 85s
12.9s
LightGBM training time
production-grade speed
LightGBM Hyperparameter Search — Validation REQM
4 combinations tested · n_estimators × num_leaves · best: n=200, leaves=50
TUNING
1,683
Best validation REQM
n=200, num_leaves=50
1,744
Worst combo
n=100, num_leaves=31
−3.5%
Improvement from tuning
vs default params
Random Forest Tuning
Validation REQM by n_estimators and max_depth
RF SEARCH
Final Test REQM — All Models
Head-to-head on unseen test data · 29,516 records
FINAL
Complete Benchmark Table
All metrics · all models · test set
TABLE
Model REQM Train Time Pred. Time vs Baseline Verdict
LightGBM €1,670.50 12.9s 0.25s −37% ✓ RECOMMENDED
Random Forest €1,727.38 85.0s 0.251s −34% Viable — slow train
Linear Regression €2,636.53 0.92s 0.08s Baseline Baseline only
Training Time (seconds)
Log scale · 236,126 training records
TRAIN
LightGBM
12.9s
Linear Reg.
0.92s
Random Forest
85.0s
Prediction Time (seconds)
29,516 test records · user-facing latency
PREDICT
Linear Reg.
0.084s
LightGBM
0.252s
Random Forest
0.251s
The Trilemma: Quality vs Train Speed vs Predict Speed
LightGBM wins 2 of 3 — and its prediction speed is acceptable for real-time use
RADAR
LightGBM
Best REQM + fastest training
The complete package
Trade-off
LightGBM slower to predict than LR
but 0.25s is app-grade performance
RF
85s training is too slow
for frequent retraining cycles
Dataset Overview
354,369 records · 16 raw features · 8 used in model
EDA
Total Records
354,369
Duplicates Removed
262
Train Set
236,126
Test + Val
29,516 each
Features Used in Model
8 retained after removing irrelevant columns
FEATURES
FeatureTypeRole
Priceint64Target variable
RegistrationYearint64Numeric predictor
Powerint64Numeric predictor
Mileageint64Numeric predictor
BrandobjectCategorical (OHE / native)
VehicleTypeobjectCategorical (OHE / native)
GearboxobjectCategorical (OHE / native)
FuelTypeobjectCategorical (OHE / native)
NotRepairedobjectCategorical (OHE / native)
DateCrawled, PostalCode…variousDropped — irrelevant
Correlation Analysis — Key Findings
Relationships between numeric features and target Price
CORRELATION
+
RegistrationYear & Power
positively correlated with Price
−
Mileage negatively correlated
higher km = lower price
✓
No multicollinearity detected
stable regression environment
MERIDIAN
Rusty Bargain · LightGBM (REQM €1,670 · 12.9s train · 0.25s predict) · 354K vehicles · Meridian Labs 2025