I dropped last_service_event_type and pickup_scheduled_at because they are written after a return is raised and provide little predictive value at dispatch time.
I de‑duplicated the 651 partner_feed rows, removing exact copies of crm rows and preventing leakage across validation folds.
I corrected October 2025 order values by dividing them by 100, fixing the paise‑to‑rupee scale error.
I validated the model on a time‑based split: train on past data, then test on the next two months, repeating four times; the hidden test set is July–September 2026.
Baseline "no‑action" accuracy is 88.7%, and the highest achievable accuracy with any threshold is 89.5%.
I used Finance’s policy return cost of ₹1,150 in the economic analysis.
I recommended calls rather than holds as the intervention, because a call prevents 35 % of returns while a hold adds 12 % cancellations.
I kept Shield members in scope; Shield customers have an 18.9 % return rate versus 9.3 % for non‑Shield.
I built the kestrel-returns-risk repo to score each order before dispatch for return likelihood (Variant A) and to recommend a pre‑dispatch call or normal dispatch.
The repository is named kestrel-returns-risk.
It lives on @Nandini15-hash's GitHub.
The project is written in Python.
It was started in 2026‑09 and last updated in 2026‑09.
The service makes no calls to any external API, so there is nothing to pay for and nothing that can fail for lack of a key.
The test suite contains 9 pytest tests.
python -m venv .venv
# Windows: .venv\Scripts\activate macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt
cd service
python -m uvicorn app:app --port 8000
I recommend calling the riskiest 20 % of orders (about 140 orders per month) before dispatch, which yields a net profit of ₹10,900 / month.