We simulated the 25K Fantasy auction 350 times.
25K Fantasy is brutal: a live auction with ~20 expert owners, no do-overs, $25K real money on the line. Honest backtesting can't just knapsack a perfect lineup at fixed prices — that's not how an auction works. So we built a real auction simulator that bids player-by-player, treats every opponent valuation as the actual 25K final price ±15% noise, and runs 50 live auctions per year. The projection model never sees year-Y data; only the auction dynamics vary.
Here's what would have actually happened.
The honest read: 25K is much harder than ODB. The field is tiny (8-24 teams), the buy-in eliminates tourists, and every owner is bidding with full information. A blind projection model averages around 10th place in a 16-team field — worse than median — and cashes about 11% of sims. But in 3 simulations out of 350, our team out-scored the actual contest champion. Those wins didn't come from year-wide auction luck — keep reading for the mechanic.
Three numbers, three different stories.
The auction sim numbers above describe one specific scenario: you run our model as a single $25K entry every year, opponents value players at market ±15% noise, no information edge over the field. That's a worst-case framing. The same MARCEL projections under a different execution model produce a dramatically different result. The honest range:
1st in 2017, 2nd in 2019, 3rd in 2024. $250K net on $175K entered. This is what happens if you got every player you wanted at the actual market price — no bidding wars. It's the projection's pure read on a year, perfectly executed.
3 outright wins across 350 sims; everything else mostly $0. Per-entry EV is −$17,500. This is what happens if you bid generically against opponents who value players at market consensus — you can't get the picks you want cheap, and you overpay for chalk you can.
Where you actually land depends on how disciplined you are. Bid hard only where projection beats market by 20%+. Pass on chalk where projection ≤ market. The bigger your information edge over the room, the closer you get to the +143% ceiling. The point is to know which players are mispriced, not to draft a fixed lineup.
MARCEL projects each player using their last 3 years of WSOP Summer fantasy points (5/4/3 weights, most recent heaviest), regressed toward 10 points based on event volume. As of this week we also blend in WSOPE results as a recent-form signal — a player who went deep at WSOPE in October is demonstrably in shape for the following Summer. WSOPE feeds the projection input only; it never counts toward scoring — 25K Fantasy grades against Summer events alone, and our backtest scoring reflects that. Signal in, no contamination out.
The randomization isn't a year-wide discount.
The intuitive question reading the table above is: did the model win 2018 and 2022 because everyone happened to bid low that year? The answer is no — the noise model doesn't work that way.
For each simulation, every opponent team generates a fresh random valuation for every individual player — uniform multiplier in [0.85, 1.15] applied to that player's actual 25K market price. So in any one sim, opponent Alice might value Negreanu at 90% of market while opponent Bob values him at 108%. The noise is per (team × player), not year-wide. There's no “cheap year” outcome where every player goes for less than they should — for every player one opponent undervalues, another opponent probably overvalues.
What CAN happen across a 16-opponent sim is that the particular player our model wants most happens to be undervalued by enough opponents simultaneously that we win the bidding cheap. With 14-24 opponents each rolling per-player noise, the probability that the second-highest valuation comes out at the low end of the range is roughly (0.5)N for the target player — small but nonzero. That's where the cash sims and the rare win sims come from.
The wins also require the player we got cheap to crush WSOP — projection skill at the cap level plus WSOP variance both have to break our way. Take 2018: in our winning sim we paid $69 for Paul Volpe (market was $61 — we actually overpaid). Then Volpe scored 313 fantasy points versus a marcel projection of 86. Shaun Deeb at $57 paid (market $52) scored 294 vs proj 74. John Hennigan at $13 scored 334. Three of our eight picks each scored 3-4x their pre-summer projection. That doesn't happen every sim — it happened in 2 of 50 in 2018.
In short: a winning sim isn't “random noise was nice that year.” It's two coincidences stacking: (a) opponent valuations happened to undervalue specific players our model targeted, so we drafted them cheap; (b) those same players happened to have monster WSOPs. With 50 sims per year × 7 years, that double-coincidence hit 3 times.
Real auction, fake auctioneers.
Every player gets a recency-weighted fantasy-points projection from the three years before this one. For 2025 we use 2022, 2023, and 2024 data — nothing else. For 2018 we use 2015, 2016, 2017. No ridge model, no peeking ahead. The projector ignores price.
Every player's baseline price is what the human owners actually paid them for at the 25K auction that year. Pulled directly from the contest's public results.
Each sim runs a full live draft. Players are nominated roughly in order of descending market price (just like the real auction). For each nomination, every team computes a max bid:
- Our team bids based on VORP — what marcel projects the player to score over the next-best available replacement. We're deterministic; same projections = same bid ceiling.
- Each opponent team samples a per-player valuation: marketPrice × Uniform(0.85, 1.15). Independent per (team, player) — there's no year-wide cheap-auction mode.
Bidding goes Vickrey-style: the high-bidder wins at one dollar above the second-highest bid. If we're outbid we lose the player and move on. At the end every team has 8 players within their $200 cap.
The lineup we end up with in each sim gets scored using that year's actual fantasy scoring rules (same buy-in multipliers, field bonuses, bracelet bonus for 2025+). We rank against the real human-submitted teams from that year. Every cash and every win you see below is against the actual contest field — no synthetic opponents in the ranking step.
For our team to out-score the actual 16-team champion, three things have to align in the same sim:
- The model's top-projected sleepers happen to fall to us in the auction (because enough opponents rolled low valuations on them).
- Those same players then crush WSOP — their actual score blows past projection (3-4x in the 2018 case).
- The expensive stars opponents outbid us on under-perform their price tag (Negreanu went for $150 in our 2018 winning sim and scored 67).
All three happened 3 times across 350 sims — a small, defensible rate. We don't claim our model crushes the elite 25K field on a single entry; we claim it has identifiable edge that occasionally compounds with auction + variance luck into an outright win.
If we'd run our model as a single entry every year — what would the bank account look like?
25K Fantasy is a $25,000-entry contest. Across 7 years × 50 simulated auctions per year (350 total entries), the model would have paid $8.75M in buy-ins and received $2.62M in payouts. The aggregate financial result is the headline below — and it's not pretty.
A single-entry strategy in 25K Fantasy, with our blind model, is a negative-EV game. Most sims return zero (we finish outside the top 4); a handful cash; very rarely we win the whole pool. The expected dollar outcome per entry is -$18K — meaning if you ran our model as your sole entry every year for 7 years (real or simulated), you'd expect to lose money. The biggest single-sim cash was $176K in 2018 (one of our 3 outright wins), but it's drowned out by 312 sims that returned zero. Single-entry, blind model = lose money. The realistic path to positive EV is multiple correlated entries + an opinionated nudge to the projections, or simply not playing 25K Fantasy when you don't have an information edge over a 20-expert field. The model is more useful here as a price-discovery tool.
| Year | Prize pool | Cashed | Won | Mean payout | EV/entry | ROI |
|---|---|---|---|---|---|---|
| 2017 | $200K | 9/50 · 18% | 0 | $5.4K | -$20K | -79% |
| 2018 | $375K | 3/50 · 6% | 2 | $9.0K | -$16K | -64% |
| 2019 | $300K | 4/50 · 8% | 0 | $4.1K | -$21K | -83% |
| 2022 | $350K | 12/50 · 24% | 1 | $19K | -$5.5K | -22% |
| 2023 | $500K | 0/50 · 0% | 0 | $0 | -$25K | -100% |
| 2024 | $475K | 8/50 · 16% | 0 | $11K | -$14K | -55% |
| 2025 | $600K | 2/50 · 4% | 0 | $3.2K | -$22K | -87% |
| All years | $2.80M | 38/350 | 3 | $7.5K | -$18K | -70% |
Each bar is one simulated auction. Height = team score; color = payout tier. The fat zinc base is the bulk of sims that finished outside the top 4 (payout: $0). Emerald = cashed (2nd-4th). Amber = outright win. The few amber bars per year are the model's rare crown moments — visible but rare enough that they don't flip the per-entry expectation positive.
Every year, three fake drafts.
Each card expands to show our worst, median, and best simulated draft for that year — the actual rosters the model would have ended up with under three different auction outcomes. Click into any year.
Players we got outbid on (6)
| Player | Went for | Our proj | Actual |
|---|---|---|---|
| Jeremy Ausmus | $159 | 134 | 86 |
| Scott Seiver | $143 | 102 | 46 |
| Yuri Dzivielevski | $136 | 114 | 95 |
| Daniel Negreanu | $120 | 91 | 269 |
| Nick Schulman | $123 | 49 | 250 |
| Jesse Lonis | $114 | 78 | 43 |
Players we got outbid on (6)
| Player | Went for | Our proj | Actual |
|---|---|---|---|
| Jeremy Ausmus | $158 | 134 | 86 |
| Scott Seiver | $141 | 102 | 46 |
| Yuri Dzivielevski | $137 | 114 | 95 |
| Daniel Negreanu | $119 | 91 | 269 |
| Nick Schulman | $123 | 49 | 250 |
| Jesse Lonis | $117 | 78 | 43 |
Players we got outbid on (6)
| Player | Went for | Our proj | Actual |
|---|---|---|---|
| Jeremy Ausmus | $160 | 134 | 86 |
| Scott Seiver | $139 | 102 | 46 |
| Yuri Dzivielevski | $131 | 114 | 95 |
| Daniel Negreanu | $124 | 91 | 269 |
| Nick Schulman | $121 | 49 | 250 |
| Jesse Lonis | $118 | 78 | 43 |
Three caveats no sharp critic should miss.
Real 25K Fantasy owners aren't bidding ±15% around the historical market price — they ARE the market. We're modeling them as a cohort of noisy market-price bidders, but the actual humans are doing complex VORP math, hedging, and counter-signaling. The model probably ran AGAINST tougher opponents than this sim produces. Treat the cash-rate numbers as an upper bound, not a floor.
25K Fantasy added a bracelet bonus in 2025. Our marcel projector doesn't know about it — it just averages past fantasy points using the rules in effect when those points were scored. A meta-aware strategy (load up on mixed-game specialists with bracelet upside) only became correct in the last cycle.
In a 15-team field, finishing 8th-10th is just bad luck given a decent process. We're running 50 sims to characterize the distribution, but in real life you only get one shot per year. The variance is brutal.
25K is the hardest WSOP fantasy contest to crack. Our model lands around the bottom half of a tiny expert field — better than random, but not a reliable cash. It's much more competitive than ODB (where we cash 43% of the time at a fraction of the field skill). That's the honest calibration: the smaller and sharper the field, the smaller our edge. The model is more useful as a price-discovery tool (which players are over/underpriced relative to projection) than as a single-entry crown machine here.