Wednesday — Half Price Tracker

The Idea

Australian supermarkets run famously aggressive half-price cycles — the same Tim Tams, laundry detergent, and vitamins drop to 50% off on a rhythm, every few weeks, always landing on a Wednesday. Several apps show you this week's specials. None of them answer the question shoppers actually have: "should I buy now, or is it about to go half price?"

Wednesday's wedge is forward prediction: search any product and see its predicted next half-price window, an honest confidence level, and the predictor's real measured track record. Free forever, no ads, no accounts, no tracking. Also live at wednesday.ernestmikhail.com.

Data Pipeline

Every day a Python pipeline (GitHub Actions + one scheduled task on a residential machine) ingests ~41,000 products across Coles and Woolworths from three complementary public sources: the open-source hotprices.org daily price dumps, Woolworths' own browse API, and Coles' published weekly catalogue as a cross-check. Half-price events are derived from price-change history, not trusted from labels — a change-point counts only when the price genuinely drops ≥48% against a windowed regular price, with a recency guard against stale data. Woolworths blocks datacenter IPs, so the live API is pulled from the residential machine while the cloud pipeline falls back to dump-derived data — and a "sticky live" guard ensures the weaker source never overwrites the stronger one.

The Predictor

Statistical, deliberately not ML. For every product with enough sale history (~18 months backfilled, ~62,000 historical half-price events), the engine computes cycle statistics — mean interval, variance, time since last sale — and forecasts the 1–3 week window when each item next goes half price, with a confidence score. Honesty is enforced structurally: under 3 observed cycles the app says "warming up" instead of guessing, confidence is capped, and expired windows display as "window passed," never as a forecast. Tested on ~91,000 past forecasts: when the model is at least 70% confident, it is right 81% of the time — and that measured track record is published inside the app on every product page.

Audited Accuracy

Half-price deals are checked daily against Coles' and Woolworths' own lists: ~99% found, ~99% correct. A 21-invariant verification suite runs at the end of every pipeline job and fails the workflow loudly if anything drifts — a green run must mean the data is actually right, not merely that the code didn't crash.

The App

React Native / Expo, TypeScript. Four tabs in a warm paper "ledger" design: this week's drops (both retailers interleaved by savings), full-catalogue search with deals ranked first, an on-device watchlist, and weekly "Half-Price Dinners" composed only from genuinely half-price items. The same product is matched across both supermarkets (~5,800 cross-store links built with precision-first matching, adversarially audited) so the app can honestly say which store is cheaper this week. Resilience is hand-rolled: transient-network retry, last-good snapshots with a Wednesday-cycle honesty guard (cached data never masquerades as the current week), and silent revalidation on foreground.

Notifications Without Accounts

There is no sign-up anywhere. Watchlists live on-device; opting into alerts registers only an anonymous random UUID, the watched product IDs, and a push token — locked behind row-level security with a single validated RPC as the only public surface. One batched digest per week ("3 of your items are half-price this week"), deduplicated at the database level, with delivery receipts swept and dead tokens pruned.

Engineering War Stories

The SKU re-key. Products were originally keyed by normalized name — which silently merged distinct same-name products. Re-keying everything to real retailer SKUs fixed a ~1% price-mismatch class and, as a bonus, made product image URLs deterministic: 100% image coverage with zero scraping.

The marketplace purge. ~73% of the Woolworths feed turned out to be third-party marketplace junk ($999 dog strollers with fake was-prices, dominating any savings-sorted list). The detector: real stockcodes are ≤7 digits, marketplace IDs are 10 — a bimodal split that purged ~53k rows and made the catalogue honest.

The Hermes release crash. A release build crashed where dev builds didn't: Hermes can't parse variable-specifier dynamic import(), which a dependency shipped in an optional telemetry loader. Fixed with a Babel plugin that rewrites them — found by reproducing the release bundle locally.

The week-rollover incident. Launch week delivered a real incident: the app's "current week" rolled the moment the first data source wrote to it, so one morning users briefly saw a fully-stocked Woolworths next to "Coles: 0 items." The fix became the project's first Architecture Decision Record — atomic week rollover: no writer may create a new week unless it arrives complete and provably fresh. The worst case is now a few hours of last week's complete data instead of an empty store.

App Review. Approved same-day, first submission — no login for reviewers to fight, honest metadata, and reviewer notes that pre-answered the affiliation and data-sourcing questions.

By the Numbers

~41,600 supermarket products tracked daily — ~2,700 half-price deals live in a typical week — ~18 months of price history, ~62,000 historical half-price events — Forecasts right 81% of the time when ≥70% confident, tested on ~91,000 past forecasts — published in-app — Deals checked daily against Coles' and Woolworths' own lists: ~99% found, ~99% correct — 130+ automated tests (Jest + pytest), 21 live-data invariants gating every pipeline run — Infrastructure cost: $0/month