Skip to content
back to projects

case study · 2026

Mercari Price Suggestion

End-to-end machine learning pipeline that predicts fair marketplace prices from listing text — capstone for the BRACU–SICIP Data Science certificate.

role

Solo — full pipeline

code
PythonPandasScikit-learnLightGBMNLP (TF-IDF)FastAPI

architecture

Data

Kaggle Mercari dataset — 1.4M+ listings

Processing

Pandas + scikit-learn sparse pipeline

Features

TF-IDF (title + description), categorical encoding

Models

Ridge baseline → LightGBM (selected)

Serving

FastAPI + serialised model & encoders

Evaluation

RMSLE on log-transformed target

The problem

Mercari is Japan's largest community-powered marketplace. Sellers constantly misprice listings: overprice and the item never sells, underprice and money is left on the table. The task — from the Kaggle Mercari Price Suggestion Challenge — is large-scale regression with NLP: predict a fair price from nothing but the listing's title, description, category, brand, condition, and shipping flag.

Evaluation metric: RMSLE (root mean squared logarithmic error), which punishes relative error — being $5 off on a $10 item is much worse than on a $200 item.

Key features

  • Complete reproducible pipeline: EDA → cleaning → feature engineering → model comparison → serving, not a single notebook
  • Sparse end-to-end: TF-IDF over title + description produces a wide sparse matrix kept sparse through every stage — the whole pipeline trains on a laptop
  • Two-model comparison with honest baselines: Ridge regression, then LightGBM on identical features and splits
  • FastAPI serving with an inference-time preprocessing module that exactly mirrors the training pipeline
  • Serialised model + encoders shipped together, so the API is one uvicorn command from a prediction

Results

ModelSplitRMSLEMAE
Ridge regression70/300.6246$14.17
Ridge regression80/200.6248$14.16
LightGBM70/300.4660$10.48
LightGBM (selected)80/200.4650$10.44

The selected LightGBM model improves 25.5% over the Ridge baseline, with a mean absolute error of $10.44 on real listing prices.

Challenges

Scale on a laptop. 1.4M listings with TF-IDF features explode into a matrix that would need a cluster if densified — one careless .toarray() call away from an out-of-memory crash. Keeping every transformer sparse-aware was the difference between "runs locally" and "needs infrastructure."

Respecting the metric. RMSLE means training on log1p(price) and inverting at prediction time; getting the transform boundary wrong silently corrupts every downstream number. The evaluation harness applies the inverse in exactly one place.

Train/serve skew. The most common failure mode of deployed ML is inference preprocessing drifting from training. Every transformation lives in one shared module imported by both paths — there is no second implementation to drift.

What I learned

  • The metric shapes the whole pipeline — target transforms, validation, and model choice all follow from RMSLE
  • Baselines earn their keep: without Ridge, "0.465" is just a number; against it, it's a defensible 25.5% lift
  • The unglamorous serving layer is the real job — mirror-the-pipeline discipline matters more than model choice