A recommender is only as good as the data decisions behind it. By the end of the day you should be able to:

• test a cleaning choice instead of trusting it;
• explain when confidence, lift and leverage pick different products, and which one a business should use;
• check a decision on data you held back before you commit to it;
• promise a number and explain the gap when reality lands;
• put it to a board as a Data Point of View: we should, because, unless.

Round	You get data up to	Scored on	New controls	Goal
1 · easy	Aug 2011	September	Clean the data, history	Most hits
2 · medium	Sep 2011	October	Rank by, minimum baskets, hide bestsellers	Most discovery hits
3 · hard	Oct 2011	November	Weight by price	Most £ of discovery items bought

Your settings carry into the next round. Each round's data opens when Victor opens that round.

• Purchase event: one product on one invoice in the scored month, for each of the 300 products in top300_products.csv. Returns (invoices starting with C) are not purchases.
• Hit: that invoice also contains at least one of the product's 3 suggestions.
• Discovery hit: a hit on a product that is not in that round's bestsellers_roundN.csv (the 10 products in the most invoices up to the cut-off). The homepage already shows those.
• £ value: for each discovery item bought, its median list price.
• Confidence = invoices with both ÷ invoices with the product. Lift = confidence ÷ share of all invoices with the suggestion. Leverage = invoices with both − what chance alone would give (invoices with A × invoices with B ÷ all invoices).
• Trade basket: an invoice with 20 or more different products. No-ID: blank CustomerID. History: the last 3 or 6 months before the cut-off, or all of it. Minimum baskets: the pair must share at least that many invoices.
• Rows we ignore everywhere: other countries, postage and fee codes (POST, DOT, M, C2, D, S, BANK CHARGES, AMAZONFEE, CRUK, PADS, B, gift vouchers), zero or negative price or quantity (except returns).

To test a setting yourself: build suggestions from data up to the end of July and score them on August, exactly as above. That's a backtest.

Part	What it rewards	Points
Value captured	Your result as a share of the best possible setting for that round. Every one of the round's settings has been scored against the real month in advance. In round 1, a plain "show the bestsellers" box counts as 0.	up to 60
Promise kept	How close your forecast was: 25 × (1 − |forecast − actual| ÷ actual), never below 0.	up to 25
Point of View	Victor scores your three lines: 5 for a clear action, 5 for evidence from your own testing, 5 for an "unless" someone could check.	up to 15

After each Drop, Victor publishes the results and your own score appears on this page. Champions: the highest total over three rounds (300 points). Tie-break: Promise kept. After round 3 each team pitches its setting, and a separate Best story vote follows.

Real: every invoice, the suggestions your setting produces, and the September, October and November results.
Assumed: a purchase of a suggested item counts as a hit. The data cannot show whether a suggestion was displayed or caused the purchase, so this measures how well your rules predict real buying, which is how shops test a recommender before running a live trial.