Sweet Instinct: the AI that rescues good stock actions
Sweet Instinct rereads the actions Sway One's filters discard. On eight shops, it finds 2.2 times more profitable actions than a gain ranking.
Sweet Instinct is the decision node of Sweet 1.3: it rereads the actions that Sway One's automatic filters have discarded and estimates, for each one, the probability that it is worth it after all. It answers in under half a second, with a quantified yes or no, within the five Sweet AIs.
Catching the good opportunities
A stock engine suggests restocks, sales and highlights. So as not to drown the retailer, filters cut what looks barely profitable. Among what is cut are genuine opportunities: Sweet Instinct is there to find them again.
It receives the action type, the category and brand, the quantities, margins, stock and recent sales. The product name, its reference and its image stay with you.
The results
On our benchmark, eight shops, four dates: 2,644 discarded actions, of which 171 would actually have earned at least €5. The whole challenge is to find them.
We compare Sweet Instinct with the simplest choice, keeping the biggest estimated gains, on the 40 best-ranked actions of each shop.
| Shop | Good actions, sorted by gain | Good actions, Sweet Instinct | Actual gain, sorted by gain | Actual gain, Sweet Instinct |
|---|---|---|---|---|
| Saddlery | 13 | 28 | €94 | €417 |
| Tennis and padel | 6 | 6 | €69 | €115 |
| Outdoor | 1 | 7 | −€2 | €89 |
| Running | 1 | 2 | −€77 | €68 |
| Menswear | 0 | 1 | −€146 | €9 |
| Womenswear | 0 | 0 | €17 | €18 |
| Textile brand | 0 | 1 | −€249 | −€76 |
| Children's brand | 0 | 1 | −€273 | −€472 |
| Total | 21 | 46 | −€566 | +€167 |
Sweet Instinct finds 2.2 times more good actions and moves the result from −€566 to +€167. It beats sorting by gain in seven shops out of eight, with an almost perfect ranking on outdoor.
Arbiter of forecasts
In Sweet 1.3, Sweet Instinct also settles between the forecasts of Horizon when the statistical model and the day-by-day forecast disagree. On the benchmark, it only had to decide 62 times, for less than $0.002 in total.
Fast and light
- 270 ms per decision at the median.
- About $0.05 for the 2,644 decisions of the benchmark.
Next step
Sweet Instinct is deliberately cautious: it only puts an action back forward from a probability of 80%. We will calibrate this threshold on the first weeks of each shop's data, so that it automatically rescues its best finds.
Frequently asked questions
Does Sweet Instinct decide in place of the retailer?
No. It puts an action back on the list of suggestions; it is always the retailer who approves.
What data is sent?
Figures, the category and the brand. The product name, its reference and its image never leave your deployment.
Why not simply keep the biggest estimated gains?
Because a high estimated gain does not tell you whether the action will work. On our benchmark, Sweet Instinct does far better than that sorting.