Decide with Data — Daloy Mobility's 90-Day Analytics Mandate

A four-round, advanced data-science-for-decisions simulation set inside Daloy Mobility, a Metro Manila ride-hailing and last-mile logistics platform (~46,000 driver-partners, ~1.9 million trips/week, PHP 24B gross bookings, PHP 4.3B net revenue, near break-even). A competitor with venture funding and superior ETA accuracy has entered the market, cancellations have climbed to 14%, and the board has handed the new Chief Analytics Officer a single flagship machine-learning mandate: one quarter (90 days), a PHP 90M budget, and a demand for measurable decision lift in production — not model accuracy. The board pays for lift, not AUC; the wrong problem, the wrong model, or the wrong threshold burns the team's credibility in a single quarter. Round 1: frame the prediction problem — score demand forecasting, ETA prediction (cleanest data, but the rival already leads there) and cancellation prediction (messier data, but it targets the 14% rate bleeding margin); define success as a decision outcome, not a model statistic, and weigh business value against data convenience. Round 2: the core accuracy-versus-cost trade-off — a simple gradient-boosted model (~82%, PHP 15M, 6 weeks, ships in the window) versus a deep spatio-temporal net on raw GPS (~89%, PHP 65M, 14 weeks, past the 90-day window) versus a black-box vendor API; plus the data strategy and an A/B-test measurement design. Round 3: set the deployment threshold that maximizes cost-weighted business value, not accuracy, knowing a missed cancellation (false negative) costs PHP 180 a trip while a needless intervention (false positive) costs PHP 40 — so the value-maximizing threshold sits well below the accuracy-maximizing 50%; design the human-in-the-loop boundary and a drift-monitoring rollback plan. Round 4: defend the initiative to the board as decision lift and pesos (not 0.86 AUC), and absorb a live shock — concept drift, a tougher rival, or a privacy regulator querying granular GPS — without over-claiming or abandoning the data-driven discipline. The math wires the concept's numbers and common errors into sticky, run-defining penalty flags: chasing the high-accuracy model that misses the deployment window yields a brilliant notebook and zero live lift; forgetting the threshold means the model changes no decision; the convenient problem cannot clear the lift bar; an accuracy-tuned threshold leaves money on the table; no drift monitoring lets the lift decay; and reporting accuracy to a board that buys lift risks the next round of funding. Final KPIs track decision lift (points on the target operating metric), net annual value (PHP M/yr), model accuracy (%) and deployment status — and the defining lesson is that a deployed, value-tuned 82% beats an undeployed 89%.