Skip to content

Strategy to the end of 2026

Findings from a working session on 28 September 2026 between the owner and Claude, over the racinglines MCP server (overview, track_record, edge_finder, sql) and the repo docs. Every number is from the database on that day unless a web source is cited. The action items live on the Roadmap (U1–U12 there); this page keeps the reasoning behind them. The pre-registered validation rules are in Paper trading and the round-16 plan in the F1 live test.

Where things stand

Update, 30 Sep 2026 (evening). Two days on, the picture below holds, with three changes. (1) The overnight VM sweep (106 jobs, 16,000 simulations) kept profile A as the best taker: +$1,389 in 2026 and +$1,432 in 2025 (PR #85); the demo taker's rebuilt paper record is +$450.46 over 23 weekends of 2025 and +$353.01 over 15 of 2026, and it paper-trades live from round 16. (2) NASCAR and MotoGP are in the app with a simple result-only model and an in-sample demo paper record on Kalshi (NASCAR 2026 Pro +$6,525.45, but every setting loses 2025; MotoGP 2026 +$31.25), so they are demos of the pipeline, not a second edge. (3) The plan for October now includes the invite-only fantasy soft launch on F1 from Thu 8 Oct, preceded by CI and a staging host (Roadmap). Per cell: Sports and exchanges.

The findings of 28 Sep:

  • The maker's +$12.4k headline is mostly Whistler. $11.3k of it is the Whistler private book, which traded against a simulated crowd. It shows the pricing works, not that there is an edge against a real market. On real tapes the maker made +$938 on Polymarket (+$285 in 2025, +$653 in 2026) and +$75 on Kalshi (+$193, −$119). The taker made +$898 on Polymarket (+$575, +$323).
  • The taker's record rests on two weekends. Britain 2026 (+$537) and Saudi Arabia 2025 (+$324) account for most of it.
  • The taker's record is a third of profile A's, by design. The params-4h search and the Edge Finder (run 1284) give A +$1,363 for 2025 and +$1,243 for 2026. The demo taker follows about a third of A's calls (profiles.DEMO_FOLLOW 0.33, Following); taking every call makes +$2,606, the sum of the two. So A's live proof should count A's own calls, not only the demo taker's followed subset.
  • Neither venue has any F1 race markets open. Polymarket has listed no race since 28 Aug. What is open is championship futures:
Market Polymarket volume Kalshi volume
Drivers' champion ~$38.0M ~$9.0M
Constructors' champion ~$4.9M ~$2.1M

Also open: "Action of the Year" (unmodeled), one standings head-to-head, and a Kalshi market on whether the Abu Dhabi GP takes place. - Maker results depend on the venue. Profile C makes +$653 on Polymarket's 2026 tape and loses $119 on Kalshi's (−$66 before fees). It fills more often on Kalshi and does worse on those fills, which suggests Kalshi's takers are better informed (Kalshi history). - Don't follow the Edge Finder's top rows. The stage-aware taker leads in-sample (+$1.9k–2.4k in 2026) but lost $316 out of sample on 2025.

The data behind it (28 Sep 2026)

201 events, 28,308 results and 1,166 athletes across two sports: F1 drivers' championships 2020–2026 (2026: 15 of 23 races done) and the UCI Downhill World Cup 2021–2026 (2026 finished at Whistler, 8 of 8). 1,286 model runs, 11,078 market links (Polymarket 7,285, Kalshi 3,793; 74 open on each), 3,869 paper positions and 8,667 strategy signals. Both live forecasts (F1 run 3, downhill run 1287) were built from code version 1ab8934-dirty, that is, from uncommitted code.

Worth checking: UCI 2023 shows 7 of 8 events completed and 2025 shows 9 of 10 (a cancelled round or a missing result each); the F1 forecast keeps run ID 3 while the downhill one is 1287, so the F1 row may be overwritten in place rather than written as a new run.

How long each strategy takes to prove

Rough estimates from the Edge Finder's 2025–26 replays of real weekends:

Avg per weekend SD Weekends to reach about 2 SE
Taker (update, A-like) ~$83 ~$262 ~40
Maker (conservative) ~$71 ~$146 ~17

The rest of 2026 has 8 races and 2027 about 24, roughly 32 weekends in all. That is enough to prove the maker. It is borderline for the taker even if every weekend runs. So:

  1. Freeze A and C now, and write down the success criteria before any live results.
  2. The maker is the main story and the taker is secondary. Retuning A partway through restarts its count.
  3. Treat each venue as a separate trial. Polymarket and Kalshi weekends both count, which helps the taker most.

Event tiers

Tier Meaning Examples
T1 A real market is listed. Paper-trade it with frozen profiles and compare with the replay F1 races on Kalshi or Polymarket, F1 championships
T2 A real market with no model. Record the tape and book, and score pricing where possible NASCAR, MotoGP, IndyCar on Kalshi
T3 No market. Run a simulated pool (a private book like Whistler's). This proves pricing only F1 weekends that neither venue lists

Listing timing: Polymarket listed 2026 races 29 days ahead until 28 Aug, and has listed none since. In 2025 Kalshi opened race markets only 2–4 days before each race, so it could still list round 16 around 30 Sep – 1 Oct.

The rest of 2026

  • Private books for rounds 16 and 17 prove pricing, not edge. The fair-price scorecard is their headline result.
  • Kalshi taker and maker signals (the open F1-9 item) come before more Polymarket work. Kalshi listed every race in 2025–26. If Polymarket stays dark, Kalshi is the only live venue.
  • A Kalshi-specific maker: tune it with its own sweep instead of porting C, and record it as a separate profile (K).
  • Championship markets are where the liquidity is. The season strategy lost $208 in 2026 because the model lagged pre-season. Paper-trade them small, as a separate sleeve, and keep them out of the maker/taker proof.
  • When race markets list anywhere, run A and C live and compare each weekend with its replay. That is the existing F1-8 process.

Week by week

Dates Event Market status Tier and action
Now – 1 Oct Kalshi recorder and signals, championship sleeve, freeze A and C – Wed 30 Sep: check both venues for R16. Thu 1 Oct 18:00 PDT: pick the scenario
2–4 Oct F1 R16, Bahrain GP at Sepang None yet; Kalshi possible If Kalshi lists: T1, A and C on Kalshi (C as a control). Otherwise T3, the private book (live/f1/2026-16.toml). Championship sleeve after the race either way. Optional: record SignalR during FP1
5–9 Oct Settle, reconcile, report on R16 – racinglines live report, the scorecard, the replay check. Build the Kalshi sprint markets
9–11 Oct F1 R17, Singapore (sprint) Same as R16 T1 on Kalshi if listed. Otherwise T3 (live/f1/2026-17.toml)
Oct weekends NASCAR Chase races (Las Vegas, Charlotte Roval, Phoenix, Talladega, Martinsville) Kalshi, liquid T2: record every tape
12–23 Oct Two-week gap – Sweep and freeze K, the cross-venue disagreement log, evaluate NASCAR data sources
23–25 Oct F1 R18, United States GP Kalshi likely T1: A, C and K, plus Polymarket if it returns
30 Oct – 1 Nov F1 R19, Mexico City Kalshi likely T1
6–8 Nov F1 R20, São Paulo, NASCAR finale (Homestead, 8 Nov), MotoGP Qatar F1 on Kalshi. NASCAR's championship market is deep. MotoGP thin F1: T1. NASCAR: T2, including the pre-race and in-race book. MotoGP: T2
9–19 Nov Mid-season review – The sizing decision. Walk-forward with R16–R20 added
20–22 Nov F1 R21, Las Vegas (Saturday race), MotoGP Portugal F1 on Kalshi F1: T1, checking the stage cut-offs for the night schedule. MotoGP: T2
27–29 Nov F1 R22, Qatar, MotoGP Valencia (finale) F1 on Kalshi, but the race could be cancelled F1: T1 once the settlement rules for cancelled races are in. MotoGP: T2
4–6 Dec F1 R23, Abu Dhabi (finale) Kalshi; championships settle T1. The championship sleeve settles. If the race doesn't happen, write up the settlement
7–31 Dec Season close – See 2027

Two dates decide the rest of the year: 30 Sep – 1 Oct, whether Kalshi lists round 16 (October is T1 or T3); and about 12 Nov, the sizing review.

Contingencies:

  • Polymarket returns: A and C go live there immediately (they were tuned on its flow), and that becomes the main record. Kalshi continues alongside.
  • Kalshi doesn't list the F1 races either: every remaining F1 weekend runs T3. The year's goal changes from proving an edge to proving the pricing: the scorecard, plus the championship sleeve and the disagreement log on the markets that are open.
  • Downhill: the 2026 season is over, so there are no live downhill events this year.

Owner decisions

Decision When Recommendation
Loss cap for R16 (max_loss = 500) Before 1 Oct Owner's call
Freeze A and C as they are Now Yes
A takes rookie Now No; keep A frozen
T1 on Kalshi vs T3 for R16 Thu 1 Oct 18:00 PDT T1 if Kalshi lists
Sizing rule ~12 Nov After 4–6 live weekends
Add NASCAR for 2027 December Only if the tapes and data hold up
Promote gridq+pretrain; reset for 2027 December Decide on the walk-forward

2027

December closes 2026 and sets up 2027:

  1. Season report: live weekends for each venue and profile against their replays, the pricing scorecard for all 8 rounds, the championship sleeve, the disagreement log.
  2. Walk-forward re-selection on 2025 and 2026 across both venues. Choose A′, C′ and K′ for 2027 and freeze them before Australia.
  3. 2027 model: the pre-season testing ingest (not seeing testing is why the 2026 championship calls were wrong), a reset decision for the second year of these regulations, and a decision on making gridq+pretrain the default.
  4. Decide whether to add NASCAR: look at the recorded Chase tapes (volume, spread, how often markets trade) and at data availability. If both hold up, build a model in Jan–Feb and paper-trade from the 2027 Daytona 500.
  5. Infrastructure: the GCP move, and the first real V2 order if validation rule 5 is met.
  6. Downhill: points validation, Elite and Junior Women data, and the ChronoRace backlog (run locally).

Through 2027: the taker takes whichever venue's price is better, and the Polymarket/Kalshi disagreements beyond fees are logged as a small edge that doesn't depend on the model. The real-money gate is 4–6 live weekends where fills match the replay, then one small V2 order, then scale up.

Other racing sports for 2027

Sport Venue and liquidity Take
NASCAR Cup Kalshi. The Cup champion market has about $4.4M of volume, and there are per-race markets Best next sport. About 36 races, public data, real volume. Ovals need a different model (pack racing, cautions)
MotoGP Kalshi. The champion market has about $20k Thin, but qualifying/sprint/race maps closely onto the F1 adapter, so it is cheap to add
IndyCar Kalshi race markets (about $6.5k for Detroit) Thin. Record it and watch
Horse racing Blocked. Churchill Downs used the Interstate Horseracing Act, and there were no Derby contracts in 2026 Skip
Downhill / cycling No exchange Keep downhill as a private-book showcase. Grand-tour listings not checked

Status 30 Sep: NASCAR and MotoGP moved from tapes to results, identified links, a simple model and a replay (the 2027 NASCAR decision still waits on the Chase tapes and a naive-baseline comparison, engine E7). IndyCar stays tape only: its results source's terms forbid automated use.

Do this now: record Kalshi's NASCAR, MotoGP and IndyCar tapes before any model exists. A 2027 backtest needs 2026 tapes, and they can only be collected while they are live.

Sources