What it prices

Counter-Strike 2 player props for kills and headshots, over maps 1 and 2 of a series. Every line those books post is priced where the inputs allow: a player’s per-map kill and headshot rates, how many rounds those maps tend to go, pre-match odds, recent player form, opponent and team rates. The live maps 1–2 model uses HLTV career, 90-day and last-20-map history. Courtside also posts single-map lines; those are priced on their own scale and are labelled as such.

Markets it does not model — deaths, assists and the rest — appear on the board with no number rather than a guess.

Why kills and headshots only

Those are the two markets with a rate behind them in the data. Pricing deaths off a kill rate produces confident nonsense, so the model does not do it.

One side, or a pass

Every line a book posts is priced on its own. The maps 1–2 model projects kills and headshots, then checks the posted line and payout. It selects a market and side only when its confidence check and a second probability check agree. A line that does not clear those checks is shown without a recommendation. Both markets keep their own projection on the board.

Because the unit is the posted line, not the player, the same player can be a pick at one book and a pass at another, or a pick on one of two matches in a day and a pass on the other.

Confidence, and why 35

Confidence scores each pick from 0 to 100, mostly on how far the projection sits from the line. For maps 1–2, a score of at least 35 and agreement from the second probability and payout check are required on the same market and side. Kills and headshots can each be recommended over or under. A pass is not recorded as a play.

A line that clears the threshold is still held back when it could belong to more than one of the player’s matches and the pick would change between them, when the player has a recorded roster change and fewer than 10 completed HLTV maps with the matchup team before the forecast (on CS2 maps 1–2; other scopes remain held), when the book posted more than one line for the player on the match and a more conservative one is quoted instead (or none, when the lines point different ways), when there is no match price or no named opponent yet (the projection uses both the match price and opponent history), when it is a PrizePicks demon line the model leans under or any goblin line (see below), or once the match has started. Those lines stay on the board, priced, so you can see them.

A line our feed last showed before the match appeared in our data is never priced against that match.

Confidence is a selection score, not a win probability. The earlier CS2 calibration did not establish different win rates within the recommended bands. That finding does not validate the current selection rule: CS2 changed again on October 10, 2026 to use the structural projection; that lane needs its own forward calibration.

Probability and EV

CS2 series props show a fixed 57% historical baseline from an older selection rule. It is not a player-specific forecast or a validated win rate for cs2-4.2. Single-map props and other esports use their own estimates; lineup leg chances come from a separate calibration.

EV is estimated return per unit at the displayed payout if that probability is accurate. It is sensitive to the price and the probability basis. A TAKE can still have negative estimated EV. Unknown payouts cannot support a precise return estimate.

Demons and goblins

PrizePicks also posts alternate lines. A demon is set above the regular line, a goblin below it, and both can only be played on More. A demon’s payout is set inside the lineup, so its individual Board row does not show a payout or EV. A demon the model leans under is shown but never recommended, because it cannot be played. A goblin pays less than a regular line and PrizePicks does not publish how much, so goblins are shown but never recommended. The Line type column on the Board says which is which.

When PrizePicks posts a regular line and a demon line for the same player, market and match, the regular line is the one recommended, because the record keeps one pick per player, market and match; the demon is shown but not recommended. If a demon pick was already frozen before the regular line appeared, that first pick stands, the same way a moved line keeps its first capture. Demon legs can appear in lineup cards, and a PrizePicks card holding one is priced on the regular power-play payout, which is less than PrizePicks pays for it.

How lineups differ

Suggested lineups use eligible recommended picks from one esport per card. CS2 leg chances use a separate calibration based on the projection’s distance from the line; other titles use their own leg estimates. These chances are not the historical baseline shown for CS2 series props. Calibration and dependence between legs affect the combined estimate.

A multi-match card never holds both of a player’s markets on the same match, because a player’s kills and headshots move together, and never two legs that could be the same match: where a line could belong to more than one of a player’s matches, every one of those matches counts as taken. Match stacks deliberately combine eligible legs from one match. The same-match adjustment is measured and used for CS2 only; for LoL, VALORANT and Dota the estimate multiplies the leg chances without a match adjustment.

Each book shows every card it can build from the games posted now: a 2-, 3- or 4-man when enough different games are up (you can still build a 5 or 6 in the slip builder). The combined chance multiplies the legs’ chances and trims the result for the card’s size, because legs on one slate miss together more often than independent chances would. A leg whose line could belong to more than one match is settled against each of them, and it is void if they do not agree.

A Play badge is a different test from Recommended: it means the lineup’s combined chance beats that book’s payout by at least 5%. The lineup record counts one Play card per book, size and day: the card as it stood when its first match was about to start.

How picks are graded

A recommended pick is frozen the moment it appears — line, side and payout — and settled against the official box score. One unit per pick, at the payout captured with it. A push lands exactly on the line and returns the stake; a void is a pick we cannot settle fairly: its own match was postponed by more than a day, produced no box score, or stopped short of the maps the line covers; the player did not play, or missed part of the maps the line covers; the line could belong to more than one of the player’s matches and those matches settle it differently; or two players in the box score share the name, so we cannot tell whose result it is; or the line was left over from the player’s earlier match. Neither is a win or a loss, and the member Results page shows the reason under each void. Win rates are judged against the rate each payout needs, never a flat 50%.

One posted line is one pick. If the same line is later matched to a different match, the first capture stands. If a book shows two lines on the same side for one player and match, only the less favourable one is frozen; if the two point opposite ways, neither is.

Original pick terms are retained. Settlement corrections and changes to what the record counts are disclosed in the dated Results ledger. When a recommendation rule changes, it applies from that date forward, and the Results page lists each change with its date.

The simulated stretch

Results before September 3, 2026 come from running the CS2 model that went live on September 3, 2026 (cs2-1.0) over historical posted lines, using only information available before each match. Those picks were not published live. They are labelled as a simulation everywhere they appear, and live capture-first tracking began September 3, 2026.

Other titles

League of Legends, VALORANT and Dota 2 are on the same board as CS2. A League or VALORANT series kills line is recommended when that title’s model takes a side, over or under, and clears its own selection threshold. Until October 13, 2026 only unders were recommended on those two titles. Single-map lines are judged by their own models and shown as Beta picks, outside the record. League and VALORANT deaths stay priced and graded, and are not recommended. VALORANT recommendations are rare by design: its chance is pulled toward the book’s line, because a projection far from the line is right less often than it looks, so most VALORANT lines show a projection and a lean without a recommendation. Each title keeps its own record from September 17, 2026. Suggested lineups are one title per card. The slip builder can mix titles, and it says so when the combined chance treats those legs as independent.

Where it is weak

• Historical performance does not establish future returns. Results shows the current sample and model changes.
• Match odds, opponent data and player history can be missing or stale; some rows then have no structural projection and cannot be recommended.
• The historical CS2 probability baseline comes from an earlier sample; it is not updated as current picks settle.
• Above the recommendation threshold the model cannot rank its own picks: a 70 is not evidenced to be better than a 40.
• Single-map lines are scored on a different scale from maps 1–2 and are not comparable.
• The cs2-4.2 selection rule needs its own forward evaluation; the old calibration cannot substitute for it.