# DOTA and COD Model Update: What Changed and Why Why DOTA needed a rebuild, what its first four days revealed, and the return of the COD model. July 13, 2026•By LCSLarry ![DOTA and COD Model Update: What Changed and Why](https://www.lcslarry.com/images/blog/dota-cod-model-update-july-2026-cover.png) **COD is back.** I benched the model after its rough June, and today it returns rebuilt, recalibrated, and with a new guardrail against its worst failure mode. Also live today: a calibration tune-up for the new DOTA model, four days after its release. Here's what changed in both and why. TL;DR - **DOTA:** four days of live data exposed a calibration imbalance — OVERs too confident, UNDERs not confident enough. Fixed. With the fix from day one, opening lines hit 65.4% and closing lines 68.6%. - **COD:** back online with rebuilt Map 1 and Map 3 models, Maps 1–3 math that accounts for correlation, and reined-in combo confidence. - **New guardrail:** map-agnostic COD Map 1, Map 3, and Maps 1–3 lines stop being bettable when a book moves the line against our pick. Favorable moves and map-conditional projections remain bettable. * * * ## DOTA: the rebuild, four days in The old DOTA model earned its retirement. It underperformed in June and had developed a clear directional bias — kill projections kept coming in too low, which tilted it toward UNDERs it shouldn't have liked. The rebuild fixed the projections. But a model is really two parts: the projection itself, and the probability scaling that decides how confident to be when that projection disagrees with the book's line. That confidence is what makes a prediction bettable. Four days of live results showed the new projections are good but the initial scaling wasn't: OVERs were overconfident and UNDERs weren't confident enough. The sample is small, but the miss was one-directional enough to act on, so I've tuned OVER confidence down and UNDER confidence up. Under the release settings, opening lines ran roughly breakeven and closing lines were very profitable. With the updated scaling applied retroactively, both jump: Opening lines +8.7 pts 56.7%65.4% 34–26 at release → 34–18 with the fix Closing lines +8.1 pts 60.5%68.6% 26–17 at release → 35–16 with the fix Small sample, obviously, but this is a clear improvement over the old model. I'm happy with where the new DOTA model is at, and I'll keep monitoring the results as the sample grows. * * * ## COD: what broke and what's different COD didn't have one clean bias like DOTA. It just drifted less accurate over time, and June is where it showed — mostly on Map 1, Maps 1–3, and combo lines. Of the three per-map component models, Map 1 was the worst offender and Map 3 was next, so both got rebuilt. Map 2 barely changed, and there's a fun reason for that below. ### The closing-line problem Let's start with the bad news: even the rebuilt models can't fully rescue June's closing lines in the backtest. Opening lines look fine; closing lines stay ugly. I attribute that to two things: 1 The books got sharper COD opening lines used to be mispriced badly enough that a bumped line still had edge left over. Not anymore — in the recent sample, once a line moves, most of the edge moves with it. 2 Map picks and bans drop about an hour before the match Bettors and books react to that information; the model doesn't use the confirmed map yet. It matters most on Map 1, where which map gets played can swing expected kills a lot. ### Combos were a math problem A combo line adds up several players' kills, and the old probability calculation treated each player as independent. They're not — they're in the same match, on the same maps, at the same pace. Ignoring that understates the variance of the combined total, which made the model more confident on combos than it had any right to be. What the old combo math assumedActual range of outcomes (correlated) Same projection, two beliefs about the spread. Pricing the narrow curve overstates how often a combo clears the line — that gap is exactly where the overconfidence came from. Maps 1–3 lines had the same disease in a milder form. The old approach (simplifying a bit) summed the Map 1, 2, and 3 projections, and since a player's maps are correlated with each other, that also understated the range of outcomes. ### What I changed - Rebuilt the Map 1 and Map 3 component models. Map 2 stays mostly untouched. - Replaced the summed Maps 1–3 calculation with one that actually models the correlation between maps. - Gave combo lines their own, more conservative probability scaling — they were overconfident even relative to regular Maps 1–3 props. - Recalibrated Map 1, 2, and 3 probabilities individually. The fun part: Map 2 got a confidence _buff_, because it was hitting better than its historical probabilities claimed. - Added a site rule for map-agnostic COD Map 1, Map 3, and Maps 1–3: if the book moves one of those lines against our pick, it's out of the bettable pool and the slip builder won't touch it. Favorable moves remain bettable. What the bumped-line rule means for you At open Hydra · Kills (Maps 1–3)O 56.5Bettable After the book bumps it Hydra · Kills (Maps 1–3)O 56.5 O 57.5BumpedNot bettable Unfavorably bumped, map-agnostic Map 1, Map 3, and Maps 1–3 rows stay visible for reference, but they lose their bettable tag and the slip builder skips them. Map-conditional projections remain bettable. As with the other bettable filters, you can still manually add a visible line to a slip. ## What happens next One more thing: I have something in the works aimed squarely at the closing-line problem. I'm keeping the details vague on purpose, but if it pans out, COD closing lines stop being the weak spot. More on that when it's ready. Where things stand: the new DOTA model has been live for four days and just got its calibration fix, and the new COD model goes live at the time of publishing this blog post. I'm confident both are real upgrades over the models they replace. From here the job is monitoring — COD closing-line performance above all, plus whether the probabilities keep hitting at the rates they claim. Every bet stays publicly graded on the site, so you can check the results right alongside me. Hope yall have a fantastic week. Larry out.