[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-ninety-eight-percent-accurate-deserves-a-methods-section":3},"\nRevenue technology's boldest number belongs to Clari, whose\nforecast product\n[markets \"98% forecast accuracy by week two\"](https://www.clari.com/products/forecast/).\nTake the number seriously, in the way serious numbers are taken,\nand read it with the method held up next to the headline. We have\nno basis to call the claim false. The problem is that, as\npublished, nobody can evaluate it at all, and the ways it resists\nevaluation are a tour of everything wrong with how this category\ntalks about accuracy.\n\nFour pieces are missing.\n\n## The four missing pieces\n\n- **The error metric.** \"98% accurate\" has no standard meaning.\n  Within a percentage band? Which band? Weighted how? Every choice\n  moves the number by points, and the sentence names none of them.\n- **The horizon.** A quarter's total, forecast in week two of a\n  thirteen-week quarter, is a prediction about a number that is\n  substantially already booked. Week-two accuracy is cheap in a\n  way week-minus-six accuracy is not, and the claim does not\n  distinguish them.\n- **The baseline.** By week two,\n  [the naive forecast is strong](/insights/the-naive-baseline-is-beating-you).\n  Simple arithmetic on the deals already in play gets most of the\n  way unaided. The achievement worth paying for is the margin over\n  that, and no baseline is mentioned.\n- **The selection.** Measured on which quarters, for which\n  customers? An average that does not say how its cases were\n  chosen invites cherry-picking, and nothing in the claim rules it\n  out.\n\nNone of these objections is exotic. They are the contents of a\nmethods section, the part of any serious empirical claim that lets\na reader check it. This category publishes the abstract without\nthe paper.\n\n## Why nobody will do this to themselves\n\nBe fair about the incentive. Revenue technology sells to revenue\nleaders in their own language, and that language rewards confident\nround numbers. A vendor who published error metrics by horizon,\nagainst a naive baseline, would be volunteering complexity into a\nmarket that punishes it, and would be\n[calibrating a number everyone else leaves vague](/insights/of-everything-called-70-percent).\nThe first mover pays a real cost. That is exactly why buyers\ncannot leave this discipline to vendors.\n\n## Four questions for procurement\n\nPut these to any revenue forecasting vendor, ours included. Which\nerror metric, exactly? Accuracy from which week, against the\nquarter's final number? What does the naive baseline score on the\nsame data, same weeks? And which quarters went into the average,\nchosen by what rule? A vendor with good answers will enjoy the\nconversation. A vendor without them will change the subject to\nworkflow.\n\nProphesee's revenue forecasts come with the methods section\nattached. They carry stated error metrics, accuracy by horizon and\nthe naive baseline alongside, computed on your own pipeline before\nyou commit to anything. Publishing the method keeps us honest, and\nit is hard to copy for the same reason it is rare; a methods\nsection can be checked. [Start here](/contact).\n",1786984936966]