Working on ranking algorithm this week
Back in town and working on some tweaks the ranking algorithm. First two tweaks implemented tonight:
Head-to-Head Breakdown
Driver ranking pages now show an interactive breakdown of how each event contributed to your Glicko-2 rating. Clicking the score change for any event opens a detailed view of every head-to-head matchup from that event — one row per competitor you shared a class with.
For each matchup you can see:
- Opponent — linked to their driver profile
- Their ARS rating — so you can see whether you were beating up or punching down
- Result — win or loss
- Their time and your time
- Margin — how many seconds separated you
- Win probability — the likelihood Glicko-2 expected you to win that matchup going in, based on the rating states at the time of the event
- Surprise — how much the actual result differed from what was expected; beating a highly-rated opponent is a bigger surprise than beating someone rated below you, and surprise is what actually moves your rating
The win probability and surprise figures are computed from replayed rating states — meaning the system re-runs history up to each event and uses the ratings as they stood at that moment, not the current ratings. For opponents who competed in different events than you, their rating at the time is approximated from their nearest known state, which the modal notes.
Results are sorted with your biggest wins at the top and biggest losses at the bottom.
Penalty Time Filtering
The ranking algorithm now automatically excludes obviously erroneous times from head-to-head competition calculations.
Some events assign penalty times to drivers who did not complete a run — for example, a 500-second placeholder for a no-show. When PAX factors are applied on top, these times can balloon further. Previously, the algorithm treated them as real competitor times, which meant a driver who legitimately ran a 98-second PAX time could be recorded as losing to a "competitor" who posted a 450-second penalty time. That comparison is meaningless and was inflating losses for drivers in those classes.
The fix: any time more than 3× the fastest legitimate time in the class is automatically excluded from ranking calculations. Using the example above, the fastest time of 98 seconds sets a threshold of 294 seconds — anything above that is dropped before head-to-head matchups are computed. The driver is still recorded as having participated in the event; they simply aren't included as a head-to-head opponent.
This threshold was chosen to be conservative enough to catch only clear outliers (penalty times typically run 4–5× the fastest time) while leaving legitimate slow runs untouched.
Other Changes?
Stay tuned for some other changes, and be sure to follow along Facebook as we'll have some discussion around potentially changing how ranking for local events could work differently than National events.