The idea is one list every Monday: the … Norwegian salmon farm sites most likely to treat for lice three to four weeks later. It is for companies that sell treatment capacity: delousing vessels, service boats, cleaner fish, treatment suppliers. It does not run live yet. I replayed it over … past weeks of public data, so you can pick any week, see the list that would have gone out, and see what happened next.
Each list ranks sites by a score: the model's estimate that the site reports a lice treatment three to four weeks later. The model behind each year was trained only on earlier years, so no list has seen its own outcome. The page opens on week … of … because its result sits on the average. In an average week the all sites list gets … of … right and the quiet sites list gets ….
Nobody needs me to sort the public map by lice, or to remember which sites treat often. So the list has to beat both. Every bar below is the share of listed sites that did report a treatment three to four weeks later, measured over all … weeks. The full width of a track stands for every listed site, so the bars are drawn to scale and can be compared across panels.
Loading the yearly figures.
Much of the skill on the all sites list is spotting sites already in a treatment rhythm. Their current supplier already knows that. The quiet sites list is the harder and more useful test, and it is right about one time in ….
Every list here comes from … to …. The last one is week … of …, because a list can only be checked once … more weeks of reports are in. Nothing on this page describes any farm today.
The model leans most on weeks since the last treatment. A site that treated lately tends to treat again, and whoever serves it today knows that already. Simply picking the sites that treated most in the last … weeks is right …% of the time, against …% for the list.
The quiet sites list drops every site that treated in the last … weeks. That is the list closer to new work. It is right …% of the time, against …% for treatment habit, …% for the lice map and …% for a random quiet site. So …% of those calls go to a site that does not treat in the window.
On the all sites list, …% of listed sites treated in the next … to … weeks and not in weeks … to …. On the quiet sites list it is …%. I count those as misses, because a call about a job three weeks out comes too late for them.
The all sites list was right …% of the time in …, …% in the year after and …% in …, which covers … weeks only. I cannot tell yet whether that drop is noise or a trend.
The score is a chance of treating, not a number of lice. I did test count predictions on the side: two weeks ahead the model's error was …% lower than repeating last week's count. Counts are not the product.
The list cannot say which supplier a site already uses, which method it will pick, or whether it is open to an offer. Scored on mechanical treatments alone it is right …% of the time. In this data about …% of the windows with a treatment include a mechanical one.
Lice counts, sea temperatures and treatments are what each farm reports every week, as published by BarentsWatch Fish Health. A treatment here means a reported one, and a late or missing report looks the same as no treatment. No operator names are used.
This is a call list built from public reports. It says nothing about whether a site should treat, and it is not veterinary or regulatory advice.
The live list does not exist yet. I am counting answers before I build it, and an honest no helps as much as a yes. Nothing is charged and no card is asked for.
I keep what you type in this form and nothing else. I will write to you only about this list, and I will not pass your details on.