Marketplace Explore 6,000+ AI-powered automations to scale your business.

Pharmaceutical Sales Forecasting

Pharmaceutical – Sales forecasting & what-if scenarios

Multi-model Streamlit forecasting application with temporal backtesting, explainability and business scenarios.

Model families benchmarked

Error metrics per backtest

A forecast factory

Models compared on robust backtests, explained, and ready for business scenarios

Project Details

Client

AbbVie

Sector

Pharmaceutical – Sales forecasting & what-if scenarios

Year

2024

Key Technologies

StreamlitDartsPyTorch LightningSHAP

Mission

Lead Senior Data Scientist - Time Series Forecasting

Sales forecasting & what-if scenarios

Hybrid assignment, France. English.

Design and development of a “ready-to-use” Streamlit application to experiment with, evaluate and compare pharma sales forecasting models on single/multi-series data. Implementation of an end-to-end workflow: CSV ingestion from the UI + sample datasets, minimal schema validation and dataset preparation for productionization. Implementation of a decision-oriented multi-model benchmark: training, robust backtesting, ranking, automatic selection of the best model, then a usable final forecast. Delivers an operational, traceable and reproducible “forecast factory”, designed for multi-model comparison and methodological robustness (backtesting, quality checks, time series safeguards). Focus on business adoption: actionable interpretability (SHAP + diagnostics), lever-oriented what-if scenarios, exports and model cards supporting validation, communication and decision-making.

Contexte & Environnement

Pharma / Sales Forecast context: forecasting volumes / patients / market shares at heterogeneous geographic granularities to make Sales decisions more reliable. Typical data: territorial identifiers (brick/territory/region), market variables (total patients), product variables (product patients), market shares, field effort proxies (interaction frequency), exogenous covariates and future scenarios.

Team

Lead Senior Data Scientist – time series

Environment

Hybrid, France – Streamlit application, Darts and PyTorch Lightning models, SHAP explainability

Objectifs Cles

1Standardize data ingestion and quality: validation, inference of a common frequency, handling of misaligned series, trimming of trailing zeros/non-informative values.

2Build a robust evaluation protocol: rolling-origin k-fold backtesting, RMSE/MAE/MAPE/sMAPE metrics, multi-series summary, ranking and selection of the “best-in-class” model.

3Productionize feature engineering: covariates (exogenous), lags, rolling features, scaling, sin/cos calendar features depending on configuration.

4Strengthen business trust and adoption: global/local explainability (SHAP), PDP/ICE diagnostics, constraint governance via configuration file,...

5Enable forward-looking decisions: what-if scenario module on future covariates, trajectory comparison, “stability” reading by segment/UGA where relevant.

6Ensure traceability & reproducibility: structured exports (forecasts, backtests, metrics), model cards, exportable figures, per-run logs, ZIP export, optional MLflow tracking.