EN

How it's computed

Positive share
Positive reviews are divided by all reviews in the smoothing window as sums, not as an average of daily shares: a day with three reviews does not weigh as much as a day with three thousand. Each day gets its own window: it grows until it holds 100 reviews, up to 29 days; with fewer than 10 reviews in the window there is no share, and the line breaks. The window never crosses the release date: different players write before and after launch.
Usual for the game
The share of positive reviews among all reviews in the 90 days before the week in question (for the scoreboard — before the last 7 days), by sums rather than an average of daily shares. Never earlier than the release date: before it, different players were writing. Fewer than 14 days of baseline after release — there is no “usual” yet, and the game runs in release mode without a magnitude.
Big events
Milestones, major updates and announcements, as labeled by the model from game news; without labeling — expansions, seasons and patches whose notes are several times longer than usual for the game. Several posts about one release within three days count as one event.
Reaction to an event
The week after an event compared with the week before: shift of the positive share with a 95% interval, growth in total and negative reviews. “Players loved it” or “mood dipped” means a shift of at least 2 pp and more than two standard errors after correcting for the game's noise; otherwise “within the game’s usual swings”. For labeled games — also the topics whose criticism shifted by 5 pp or more. A rebound is a significant shift within two weeks after a shift the other way, if the share came back to its earlier level (within 3 pp or half the margin); if it went further, it is a rise or a dip “after”, with what came out in between. For launches, seasons, expansions and major updates the outcome is read from the event and is never a rebound.
Shifts
Days when the 7-day average share after differs markedly from the 7 days before. A shift next to a serious event becomes part of it; a shift without an event stays in the log only if the shift exceeds three standard errors adjusted for the game's noise — stricter than for an event: the detector picks such a day itself, and a softer test would find shifts in pure noise. Timing is a hint, not proof of cause.
Game noise
How much more the daily positive share jumps around than pure chance would allow: big games have their own “moods of the day” — weekends, streams, waves of coverage. Every estimate divides the evidence by this noise, otherwise for a game with thousands of daily reviews any trifle would look significant.
Weekly shake
The last 7 days against the 90 days before. Mood is the positive share and its shift from the baseline, adjusted for the game's noise; activity is reviews per day against the usual level. The mood shift becomes the magnitude M. A surge in reviews without a mood shift does not raise the magnitude and is labeled separately — “reviews ×3”. “Shaking now” in the filter and the top of the “by shake” order mean M2 or higher or three times the usual reviews; next come M1 or 1.8 times the usual reviews.
Magnitude M0-M5
The strength of the week's mood shift — the lower of two steps: by significance (from 2, 3, 5, 8 and 13 standard errors adjusted for the game's noise) and by size (from 2, 3, 5, 8 and 12 pp). So a shift counts only if it is both not random and noticeable: for a game with thousands of reviews any trifle is significant, for a small one even 10 pp can be noise. M0 — calm, M1 — tremor, M2-M3 — shaking, M4-M5 — quake. M compares the week with the previous three months, so slow mood drift raises it too: it means “not as usual”, not necessarily “something happened”. What happened is in the shift log.
Words
Share of reviews with a word over the period compared with the 90 days before it, with add-one smoothing so rare words don't shoot up. English and Russian stop words are dropped; other languages are left out.
Model labeling and analyses
The model tags a review with catalog topics marked “praised/criticized”. A sample is labeled — up to 30 reviews a day, more on shaking days — so topic shares are of labeled reviews. Event analyses and monthly digests are written by the model from ready numbers in both languages at once, and code checks every number, every supporting review and that both languages say the same thing; anything that fails isn't shown.

↑↓ to choose, Enter to open, Esc to close