Games selected by thousands of players in the last 24 hours: How to count, filter and rank without distortion

1) What exactly we consider "in the last 24 hours"

Window: sliding T − 24h... T with a fixed time window (one for the entire project).
Segments: mobile/desktop, new/returning audience, key GEO (if relevant).
Deduplication: one user = one account (take into account device-fingerprint/anti-fraud), combine skins of one game into an active version.

2) Minimum eligibility thresholds

The game falls into the pool only if at least one of the following is completed in 24h:

Unique Starts ≥ 1,000 * or Sessions ≥ 4,000 (adjust the scale for the project).

And simultaneously:

Crash/Error Rate ≤ 0.5% of sessions.
TTI (web) ≤ 6 s at the "cold" start.
No traffic anomalies (see item 5).

3) Kernel metrics (collect "as is" and smoothed)

Fraction smoothing (beta smoothing):

$$
\ hat p =\frac {ext {successes} +\alpha} {ext {trials} +\alpha +\beta}
$$

Real time averages (EMA by the hour):

Weight of hour $ h $ back: $ w _ h = 0 {,} 5 ^ {h/au} $, half-life $au = 6 $ h.
Smoothed value: $\overline x =\sum (w_h x_h )/\sum w_h$.

4) TrendLift over 24 hours

5) Anti-anomalies and data cleansing

We cut off candidates if at least one is found:

Burst from one traffic source/GEO (> 70% of all starts) without confirmation in the next 6-12 hours.
Suspicious repetitions (many accounts with one device/IP), obvious farm patterns of auto-spins.
Duplicates of "skins" of the same game in different categories (we leave the most complete version).
Distortion in UX (forced auto-start without clicking, "stuck" tutorials).
Technical problems: Crash> 0.5%, TTI> 6 s, frequent asset reboots.

6) Integral scoring "Choice 24 hours"

We normalize the metrics by percentiles inside the 24h window (winsorize 1-2%) and consider:

extbf{Score}_{24h} = 0{,}14\cdot CR + 0{,}14\cdot ASL + 0{,}08\cdot SPS + 0{,}08\cdot (BER	MFR) + 0{,}12\cdot Fav + 0{,}10\cdot Reopen + 0{,}10\cdot (PW100	NPS) + 0{,}08\cdot UStarts_{norm} + 0{,}16\cdot TrendLift - ext{Penalties}
$$

7) Threshold for admission to the final showcase

The game is shown in the block if at the same time:

Fulfilled the thresholds of paragraph 2 and passed anti-anomalies.
CR (smooth) ≥ 38%, ASL/SPS ≥ window median.
BER ≈ 1/85-1/140 * or MFR = event every 8-25 spins.
Fav/Add ≥ 10%, Reopen ≥ 30% (both clear) .
PW100 ≥ 10 * or NPS ≥ 30%.
Score\_ {24h} in the top quartile among candidates.

8) Window balance (so that the list does not "go sour")

9) Game card template

Fact string:

* Mechanics· Volatility· Rhythm: event every N-M spins· BER (or MFR)· ASL/SPS ≥ medians· Fav/Add X-Y%· Reopen Z%· Trend ↑ (TrendLift) *

Short extension (3-4 points):

what the base "holds" (mini-events/random mods),
how the bonus feels (frequency/retriggers),
for whom (beginners/experienced),
technical advantages (fast start, stability).

10) Why exactly these games end up in the "24-hour pick"

Typical features:

1. Live rhythm - events every 8-25 spins, few "quiet" episodes.
2. Available bonus or offsetting MFR - the player quickly receives reinforcements.
3. Pure UX - TTI ≤ 6 s, animation skip, large buttons, correct return from the background.
4. Strong behavioral reaction - the growth of Fav/Add, high Reopen in the first hours.
5. Robustness - metrics hold after smoothing and clearing traffic.

11) Block publication checklist

12) Red flags (exclude from the showcase)

13) Responsibly