Investment Memo · Confidential · 10 June 2026
A read on a small, high-variance, information-advantaged betting partnership. Every number is sourced to our own analysis. Where it's unproven, it says so.
This is a $2,500 bankroll bet into a member-member golf Calcutta — a live auction where you buy teams and collect if your team wins or places in its 6-team flight. Five years of results say flight outcomes in this 9-hole net-better-ball format are statistically indistinguishable from a lottery (year-over-year skill correlation r = −0.12, CI [−0.41, +0.17], skill share of variance ~0%).
The single-Wednesday expected return is not a confident positive number — it is near break-even unless our private information edge is real, in which case the model says +8% to +24% after disciplined sizing. The ask is ~$1,200–$1,500 deployed (not the full $2,500), for an asymmetric, low-downside, repeatable bet whose asset is a proprietary information + execution pipeline, not the outcome of one auction. If you want a sure thing or a scalable fund, this is not it.
| # | The hard fact | The number |
|---|---|---|
| B1 | Outcomes are near-random. Finishing well one year does not predict the next. | team r = −0.12, CI [−0.41, +0.17]; skill share ~0% |
| B2 | No price rule beats the field with confidence. Every positive backtest strategy's CI includes large losses. | mid-price +4.0%, CI [−44.7%, +57.0%] |
| B3 | Favorites are a coin flip that flipped. The "market predicts winners" thesis rests on one anomalous year. | −38% / −37% (2024) vs +18% / +12% (2025) |
| B4 | The entire edge rests on ~2 sandbagger teams. 64% of model edge from Wood+Estes & Wright+Gadsby. | $390 / $611 |
| B5 | That sandbagger edge is statistically unproven. Best signal, least trustworthy: 7 picks, 2 years, 2 wins dominate. | +76.7%, CI [−45.0%, +192.1%], n=7 |
| B6 | ~1/3 of random portfolios made money. A two-year positive result is not evidence of skill. | 32% profitable |
| B7 | The buyback is the owner's option against us (adverse selection): reclaim winners, dump losers. | +$308 → −$11 mean ($319 swing) |
| B8 | Winner's curse is real. We win auctions on teams we like because we value them above the room. | −$600 EV → break-even |
| B9 | Risk of ruin is non-trivial — even assuming our model is right. | P(lose) 38.6%; P(lose ½) 20.1%; P(wipeout) 6.7% |
| B10 | It doesn't scale. A single ~$84k pool, capped bids ($2,000), one night a year. | total pot $84,100 |
Under a pessimistic-but-plausible stack (pay our ceiling + pot 15% light + sandbaggers convert at 75% + no buyback relief), the model's own red-team puts EV at −$561 (−18% ROI). Under fully efficient pricing it is −$250 on $2,500.
The thesis is not "we predict winners." Section 2 proves we largely can't. It's three narrower claims:
Four sources, stacked by trust: (1) a GHIN-driven Monte-Carlo valuation — 60k hole-by-hole sims/flight, audited as correct and unbiased (Gelman-Rubin R-hat = 1.000); (2) five years of real results (2021–25, 466 rows) — the only signal that survives an honest out-of-sample test, and orthogonal to handicap and to the model (r = 0.05); (3) a Cap Patrol form/clutch overlay, used as a cross-check only; and (4) insider human reads — the genuinely proprietary input. A partner personally watched a flagged sandbagger (Williford) shoot 73 in a four-club tournament — near-scratch ability behind ~8–9 strokes of phantom handicap GHIN cannot see. Copeland is a known sandbagger who won Flight 5 in 2025 — result-confirmed. No other bidder combines a calibrated simulator, 5 years of normalized results, a form/clutch overlay, and decades of inside knowledge. Caveat: it moved only ~2 teams enough to matter, on n=7. Thin and unproven at scale.
In the forward Monte-Carlo, winning bidding wars (paying our ceiling) cuts E[ROI] from +21.8% to +9.1% — a ~13-point swing; the red-team prices the same effect at −$600 EV. The single most valuable behavior is not bidding when the price crosses our number — enforced live by a tool that re-prices every unsold team after each sale.
A research-grade flight simulator (60k-quality precision at ~15–20k sims; uncertainty bands on every P(win)); a normalized 5-year results database with collision handling; a Cap Patrol overlay distilled to two real signals; and a live auction tool tracking 20 flight pots and enforcing walk-away limits. That cost is sunk and improves every year as more results land — at 5+ years the repeatability CI tightens enough to settle whether any skill signal exists. The single-year edge is thin.
Bankroll $2,500. Payout within each flight: winner 70%, runner-up 30%, with 10% of the pot directed to the shootout. Buyback — team can buy back half; planning assumption is adverse selection (own 50% of winners, 100% of losers). Pot $84,100 is a projection, not a fit to 2026 prices; pot risk is asymmetric against us (a 20%-light pot zeros the edge).
| Flt | Team | P(win) | Fair $ | Est $ | Edge $ | Why |
|---|---|---|---|---|---|---|
| 9 | Wood + Estes | 43% | 1,424 | 700 | +724 | Sandbagger and won Fl 7 in 2025 — two independent signals agree |
| 12 | Wright + Gadsby | 25% | 781 | 500 | +281 | Flagged sandbagger; model bargain |
| 5 | Wright + Copeland | 20% | 1,089 | 900 | +189 | Copeland sandbag result-confirmed (won Fl 5, 2025) |
| 2 | Downey + Shearer | 23% | 1,545 | 1,300 | +245 | Won Fl 2 in 2025; 5-yr consistent; clean names |
Hard rules (each counters a quantified failure mode): buy only $200+ below est-price (defends the −$600 winner's curse); avoid favorites/expensive flights (lost 38% in 2024); demand the result, not the reputation (Williford finished 5th in 2025 despite scratch ball-striking); price the buyback as the owner's option against you.
| Scenario | Assumes | Slate E[ROI] | E[P&L] | P(profit) | P(lose ½) |
|---|---|---|---|---|---|
| Sharp | Model probabilities roughly right | +18% to +37% | +$116 to +$514 | 51–59% | 16–25% |
| Shrunk | Probabilities pulled toward uniform | +8% to +19% | +$59 to +$262 | 45–52% | 22–32% |
| Lottery | Room prices efficiently; no edge | −4% to −14% | −$30 to −$258 | 35–44% | 29–44% |
Downside: even in the model-optimistic case, P(lose money) = 38.6% and P(lose ≥½ stake) = 20.1%. Under the pessimistic stack, central EV is −$561.
Why $1,200–$1,500, not $2,500: the 6-team slate doesn't fit the budget (half-back ~$3,021 = 121%; adverse ~$4,293 = 172%). The red-team's explicit instruction: cut stake to $1,000–$1,500. Deploy ~60% (~$1,500) on a 4–5 team value-tilted slate; hold ~40% dry powder.
The bet is small, asymmetric, disciplined, information-advantaged, and repeatable. Low-downside (no catastrophic-loss path; loss bounded by a stake sized to lose); asymmetric (top team +104% modeled edge; slate upside +28% to +37% vs a bounded downside); information-advantaged (eyewitness reads + 5-yr results no other bidder combines); disciplined (walk-away worth ~13 ROI points, enforced by tooling); and repeatable. The stake is in a repeatable, improving edge-finding operation in an information-poor market: size it small, play it many times, let the process and the data decide.
Q1. If you can't predict winners, why isn't this just gambling?
We arbitrage prices, not outcomes. Outcomes are near-random, so the edge is buying below the censored-corrected fair-value curve using two sources the room lacks (eyewitness sandbagger reads + 5-yr results). Caveat: unproven at scale (B5). It is advantaged gambling, sized to survive being wrong.
Q2. Your edge is 7 data points. That's noise.
+76.7% but n=7, CI [−45%, +192%], 2 wins dominate. It is a lead to track forward. That is why we bet $1,200–$1,500, and why the durable thesis is the pipeline that converts more years into a real answer, not this year's 7 points.
Q3. What's your real, durable moat?
Ranked: (1) proprietary human intel — you can't buy "I watched him shoot 73 in a four-club event"; (2) the integrated pipeline no casual bidder replicates; (3) walk-away discipline worth ~13 ROI points. The model alone is not a moat — decades-experienced members price nearly as well.
Q4. Why should I trust the sandbagger reads?
Only where a result backs them. Copeland won Fl 5 in 2025; Wood+Estes won Fl 7 in 2025 — confirmed. Distrust the reputation-only ones: Williford, the headline sandbagger, finished 5th in 2025. We demand the outcome, not the game.
Q5. The buyback protects your downside, right?
No — it's the owner's option against us: they reclaim winners (we keep 50%), dump losers (we keep 100%). A $319 swing from +$308 to −$11 mean. We price it as a cost, not protection.
Q6. What if the room is efficient and your model is just wrong?
Then EV is −$250 on $2,500. That is the default, not a tail; the absence of persistence suggests the model is partly wrong. Defense: bet size, walk-away discipline, dry powder. We do not bet the budget on the model being right.
Q7. What's the scale — this is a $2,500 pool?
Yes. Total pot $84,100, capped bids, once a year. This is not a scalable fund. What scales is the method: the pipeline applies to any info-asymmetric small-pool auction and improves with each year of data.
Q8. Give me the one-line EV.
Single-year ROI between −18% (pessimistic) and +24% (model-optimistic), center near break-even-to-slightly-positive only if the information edge is real, with a ~20% chance of losing half the stake regardless.
| Term | Detail |
|---|---|
| Stake requested | $1,200–$1,500 working capital (~50–60% of the $2,500 bankroll); rest held as dry powder |
| What you're buying | A stake in a repeatable edge-finding operation — proprietary intel + audited pipeline + walk-away discipline — not one auction's outcome |
| Return range (1 yr) | −18% (pessimistic) to +24% (model-optimistic); center ~break-even to modestly positive iff the info edge is real |
| Max downside | Bounded by the sized stake; ~20% chance of losing ~half in a year even when the model is right. No catastrophic-loss path. |
| The real return | Compounding — each year of results sharpens the skill estimate and the pipeline. The asset is the operation, not the Wednesday. |