What worked, what didn't, and the new Anchor strategy

A retrospective over every resolved Kalmari pick, and the strategy we built from it

2026-09-08 · 552 resolved picks (93 traded, 459 passed on but scored) · 172 research reports · strategies Keel, Sentinel, Sextant, Tracer, Ballast, Tide

One pattern dominated everything else: theses that argued something would not happen by a deadline beat the market. Theses that argued a number would get somewhere lost, badly. The new Anchor strategy only takes the side the record supports.

1. The record in one table

Every resolved pick since June, from every strategy, including the picks each run logged but chose not to trade (they are scored on the real outcome exactly like traded picks). "vs market" is the win rate minus the entry price, so it measures whether the pick beat what the market already knew. ROI is profit divided by cost.

+9 pts

NO picks on event markets vs market (n=112)

−12 pts

YES picks, all markets (n=305)

−25 pts

YES picks on data-series markets (n=155)

85%

Win rate, NO on events at 60–89¢ (n=40)

Cut	n	Win rate	vs market	ROI
Everything	532	55%	−3.8 pts	−6.5%
NO side, event market	112	77%	+9.1 pts	+13%
NO side, event market, 60–89¢ entry	40	85%	+8.2 pts	+11%
NO side, data-series market	115	69%	+5.0 pts	+8%
YES side, event market	158	54%	+3.3 pts	+3%
YES side, data-series market	155	28%	−24.7 pts	−46%

"Event markets" ask whether a person or institution does something by a date: a departure, a bill becoming law, a confirmation, a launch, a deal. "Data-series markets" ask where a number lands: gas prices, CPI, approval averages, weekly ship transits, leaderboard scores. The side effect held inside nearly every thesis type, in traded and passed-on picks alike, and in every week from June to August. It is not a price artifact: NO picks at 60–80¢ beat the market by 11 points while YES picks in the same band trailed it by 23.

2. What worked

The winning theses share one shape. For YES to pay, someone must act or a process must finish by a date, and the thesis names a structural reason it will not happen in time. None of them required forecasting a moving number.

The calendar is closed

"The House is gone July 23 to Aug 31 and the Senate hasn't passed any version — there is no physical path to a signed law before Sep 1." (Crypto market-structure bill, NO at 83¢, won.) "The 10-day presentment clock can't make the bill law by Jul 1." (Housing bill, NO at 70¢, won.)

The status quo holds

Hegseth, Patel, Alito and Pirro all staying in post; Mar-a-Lago seasonally closed so zero trips; Gemini 3.5 Pro absent from Google's own catalog. Status-quo theses won 86% of the time across 43 picks.

The resolver already agrees

"Issuance of the Trump-portrait passport began Jul 6 under saturation press coverage" (YES at 74¢, won). "The Brazil Section 301 memorandum was signed Jul 15 and is in the Federal Register" (tariffs YES at 75¢, won). The event had happened; the market was behind the resolver's publication lag.

Fading a headline where the resolver is mechanical

"The July hike premium was built during the Hormuz oil spike; oil has since crashed 40% but HOLD has not repriced" (Fed hold, YES at 77¢, won). Seven of eight overreaction fades won.

3. What did not work

The losing theses extrapolate a moving number over a short horizon, bet that a process finishes on time, or argue the market has misread the rules. In about 20 of the 31 traded losses, the write-up's own "cleanest way to lose" is exactly what happened, and that tail was almost always "the current trend continues".

Trend extrapolation on a noisy series

"AAA is $4.06 and rising 2¢/day; the rung needs only +10¢ in 9 days" (gas, YES at 83¢, lost: it stalled at $4.10). "RCP average is 40.9; below 40.2 needs a −0.7 move in 4 days" (approval, NO at 94¢, lost to two shock polls). "PortWatch dailies have been 9–12 for three weeks; persistence puts this week near 72" (Hormuz transits, YES at 70¢, lost). Trend theses won 33% of the time and returned −44%.

YES on a macro threshold

"Cleveland Fed nowcasts June core at 2.85%; a NO needs a never-in-this-cycle 0.0% core print" (lost: core printed 0.0%). YES picks on CPI, gas and payroll thresholds went 1 for 16. The NO side of the same markets went 10 for 11.

Betting a process completes on time

"Two committee windows before the deadline, and leadership wants its AG seated before recess" (Blanche, YES at 62¢, lost when the markup was cancelled). "A failed cloture vote still resolves YES and leadership says the vote happens regardless" (CLARITY Act, YES at 61¢, lost: cloture was never filed). Legislative calendars slip; theses that needed them not to slip lost.

"The market misreads the rules"

Every YES thesis whose edge was a reading of the resolution criteria lost: 0 for 6. A rules reading can support a status-quo NO; it was never the edge on its own.

Lagged resolvers

The largest dollar losses came from betting YES against a resolver that publishes with a lag while specialists traded off live data. As one report put it after the fact: "the 'mispricing' was the resolver's publication lag. The crowd wasn't lazy; I was six days blind."

Three smaller findings

Confidence hurt. "High" conviction picks won 51% of the time; "medium" won 70%. Assertive language ("lock", "mechanical", "cannot") in a thesis predicted a worse outcome, not a better one.
Citing data did not help. The most heavily sourced theses were the arithmetic ones that lost. Structural arguments that needed no nowcast did better with fewer citations.
Short horizons lost. Picks with under 7 days to close lost on both sides. Picks with 30–60 days to close won 92% of the time.

4. The new strategy: Anchor

Anchor keeps everything that already worked in the research pipeline (the series allowlist, committing a probability before looking at the price, the store of settled precedents, one leg per cluster, the liquidity floor) and adds the rules the record demanded. It launched on 2026-09-08 as an experimental strategy: its picks appear in the Kalmari feed with the 🧪 tag, members can follow it and copy by hand, and it does not trade automatically or count toward the public track record until it has earned that on its own results.

Side rule. NO is the default side. A YES pick is allowed only if the outcome has already happened and the market is behind the resolver, or a dated step of a process already in motion is on an official calendar. Every other YES thesis is a logged reject.

Event markets only. Price levels, averages, poll averages and cumulative counts (transits, executive orders, trips) are out of the universe, and any thesis that reduces to "a number crosses a line" is rejected, on either side.

The continuation-tail test. Each pick names the single cleanest way it loses. If that tail is the current trend continuing, the pick is rejected. Only tails that require a new discrete event are acceptable.

No rules-reading edges. "The market misreads the resolution criteria" is a hard reject.

Horizon and band. At least 7 days to close. Entry between 60¢ and 89¢ for the side bought (half size above 85¢); nothing above 90¢ and nothing below 60¢. No coin-flip lane, no lottery lane.

Flat sizing. Conviction is recorded for the scoreboard but no longer changes position size. Every pick caps at 10% of capital.

Backtest of the rule set on the same 532 picks

Rule set	n	Win rate	vs market	ROI
Everything the old strategies produced	532	55%	−3.8 pts	−6.5%
Anchor rules (NO, event market, 60–89¢, ≥7 days)	29	83%	+5.4 pts	+7.0%
Everything Anchor would have excluded	503	53%	−4.3 pts	−7.5%

The rule set was positive in June, July and August taken separately. Two honest caveats: it keeps only about one pick per run day, and the edge after slippage is closer to +5% than +10%. This is a small, steady edge, not a big one. Expect many weeks with zero picks and the capital held in cash.

The first run

Anchor's first live run on 2026-09-08 swept 672 open markets in its universe, found 49 closing inside 45 days, and made one pick: NO at 71¢ on "Anthropic confirms an IPO before Oct 17", mechanism "no vehicle" (only a confidential draft S-1 exists, the public filing is reported for late September and the roadshow for mid-October, so the SEC's 15-day rule plus a normal roadshow lands pricing after the deadline). The other 48 markets failed a mechanical cut before any thesis was written. Full report.

Pre-committed checkpoint. At 40 counted picks, if Anchor's record is below the market it stops and we write that up. If it clears the market by two standard errors, it graduates toward the flagship feed. Auto-copy is a separate, later decision.

5. What this means for the flagship

Keel v4, the current flagship version, already bans the biggest losers found here (continuous-number theses, sub-40¢ tails, 90¢-plus favorites, multi-leg clusters). What it lacked was the side rule. Rather than rewrite Keel and blend a new book into its older record, Anchor runs as its own strategy with its own scoreboard. Keel continues on its schedule. Both now run on the newest research model.