Forecasting pipelines that ran in production, not a notebook

Ad-inventory forecasting at an ad-tech company — model comparison, feature engineering, and deployment for live inventory planning.

Role

Data Scientist Intern — owned the forecasting pipeline

Goal

Accurate ad-inventory forecasts for live planning

Key outcome

12% MAPE, deployed to production

Context

MiQ Digital · Bangalore, India

12%

MAPE — accurate enough to drive live inventory planning

model families compared head-to-head: XGBoost · GBR · SARIMAX

LIVE

deployed to production — real decisions ran on these forecasts

$ cat problem.md

Inventory planning is only as good as its forecast

Ad inventory swings with seasonality, holidays, and campaign cycles. Underforecast and you leave money on the table; overforecast and you overcommit. The business needed forecasts it could plan against daily.

$ ./approach.sh

01Built production pipelines comparing XGBoost, GBR, and SARIMAX on equal footing.

02Engineered seasonality and holiday-calendar features that carried most of the accuracy gain.

03Reached 12% MAPE and deployed for live inventory planning.

$ cat results.json

12% MAPE on ad-inventory forecasts
Deployed to production — used for live planning decisions
Industry experience: real data, real stakeholders, real deadlines

$ grep -r skills

time-series forecastingXGBoostSARIMAXfeature engineeringproduction deploymentstakeholder delivery