Kalshi portfolio plan
Friday, October 23, 2026, evening. Where the trading money should sit, line by line, and what each line has to prove before it gets more. Written by an agent that judged the desks and built none of them. Research only: nothing here is an order.
The answer in five lines
- More than one trade is already working. Eight live weather picks went in tonight, next to the perps engine's gold short.
- No line has yet shown it makes money with real cash. The best-supported is the weather rule, at an honest +0 to +5¢ a contract after fees. Everything else is negative, unproven, or still on paper.
- Evidence is the limit, not money. Weather's order books could take 10 to 20 times today's size, and the perps engine is short of signals, not room. Adding money now adds fees, not profit.
- The fastest honest way to put more money to work is the weather ladder. It steps up after 30 settled picks, about 4 to 5 nights. Build the step-up code and move cash in now, so it can climb the day it earns it.
- Options get no new money. Hand-trading perps has no rule behind it and is the line that lost money this week; it should get a written daily loss limit, not more money.
Worst automated day
about 6.5%
of the Kalshi money: perps engine $10 + weather $9
Biggest single-line daily loss
$10
except hand trades, which have no cap yet
Weather, best estimate
+3¢
a contract after fees (range +0 to +5¢)
Perps engine, best estimate
−0.03 R
an idea (range −0.25 to +0.12 R); 1 R = $5
Where the money sits, now and proposed
Now
Proposed, once weather earns stage 2 (about Oct 28 to 30)
Perps engineWeather (predictions)Hand-traded perps
Options stay at no new money. Balances are shown as shares, not dollars.
| Pot | Proposal | Size and stage | Loss limits | Why |
|---|---|---|---|---|
| Perps engine | unchanged | $5 risk a trade, at most 3 open, 5x notional cap | $10 a day; the plan's automatic step-downs | It isn't using what it has. Its signals bind first; more money would only raise the notional cap, with no record to justify it. |
| Weather | grow as it earns | Stage 1 now: 1 contract a pick, $9 a night. Stage 2 when earned: 2 contracts, $18 a night | $5 a day at stage 1 ($10 at stage 2); back to paper on a $4-per-contract drawdown or 5 losses in a row; any glitch means paper | The only line with a positive record behind it, and the order books can take far more. |
| Rain, Miami, mentions (other prediction desks) | paper until gated | — | — | Only 6 to 7 days of forward record each. |
| Hand-traded perps | shrink; move some to weather | The trader's own | Suggested: a written $10-a-day hand limit | The line that lost money this week, with no rule. It no longer blocks the weather desk. |
| Options | no new money | Paper desks only | — | Independent review found no edge yet. |
Every line, with its record
"Per pick" is after fees at the price actually paid. Ranges are rough 95% ranges. R is one unit of planned risk ($5 on the perps engine).
| Line | Record so far | After-fee result | Room to grow | Verdict |
|---|---|---|---|---|
| Weather: day-ahead temperature favourite live, stage 1 | Backtest +8.2¢ (248 picks, but the second half only +1.8¢); out of sample +7.8¢ (52); paper +2.5¢ (48, 73% won); earlier live picks 9 of 13 won, about flat | +0 to +5¢ a contract, best guess +3¢ | About $100–150 a night across 24 cities, highs and lows; the order books hold 16–32 contracts at the ask | grow it |
| Perps engine (trend with funding; fade a stretched day) live, $5 risk | Replay Aug 28–Oct 22: 113 ideas, −0.07 R each. Paper Oct 21–23: 8 closed, +0.53 R (all shorts in a falling market). Live: 1 closed trade, flat; gold short open | about −0.03 R an idea (−0.25 to +0.12) | Signals bind first: 4 ideas in 19½ hours | hold size |
| 10 AM gold dip paper | Study: 30 weekdays, gold −7.3 bp from 10:00 to 10:05 AM; one of 36 windows tested. Paper: 1 day, +0.67 R | Unknown; study says about +6 bp net | One 5-minute trade a weekday | gate on Nov 20 |
| Funding carry | No record. Unhedged it's a direction bet; hedged it's about 6–7% a year | About $1–2 a month at this size | — | not worth building |
| Rain long-shot No paper | Held-back test +3.94¢ (35, all won); discovery +0.71¢ (345); forward paper 11 of 11, +3.45¢ | +3.45¢, but one loss erases about 25 wins | About 2 picks a day | gate ~Nov 6 |
| Miami hourly temperature paper | 35 settled, 91% won | +2.37¢ (−4.3 to +9.0) | Moves with weather; counts inside its cap | check Oct 30 |
| Mentions long-shot No paper | 44 settled, 70% won | −4.5¢ (−11.3 to +2.3) | — | losing |
| Gas and Treasury-yield favourites paper | Gas favourite 154 settled; gas shoulder 56; yields 7 | Gas −5.6¢; shoulder −0.8¢; yields −17¢ | — | losing |
| Sports vs Pinnacle paper | 3 picks in 379 games: Kalshi's game prices already match Pinnacle's | — | Two new maker-side rules going on paper | paper |
| Options desks | Grade B: 71 closed, a coin flip on direction, about half the loss is spread and fees. Grade A: positive, but all from two puts. PotterBox: 0 of 49 rows tradable | No edge shown | — | no money |
| Hand-traded perps | No rule; both days we can see lost money | No record | — | cap it |
The three changes that put money to work soonest
1. Build the weather ladder now
- Why: the longest positive record, and the books take 16–32 contracts where we use 1–8. Stages 2 to 4 still need code and about $45 of predictions cash.
- Rule (fixed): step up after 30 settled live picks at a stage with at least +5¢ a contract a pick and no glitch; step down on a $4-per-contract drawdown or 5 losses in a row; any glitch means paper. Each step up is signed by the trader. Picks that fail a check are never sent.
- Earliest effect: stage 2 around Oct 28–30. Then $9 → $18 → $36 → $72 a night as each 30 picks earns it.
2. Rain long-shot No, behind a written gate
- Why: the only paper line besides weather that passed its held-back test; forward paper is 11 for 11. One loss wipes out about 25 wins, so the gate is about losses.
- Gate: at least 30 forward picks over 15+ days, at least +2¢ a contract, at most 1 loss. Then 1 contract a pick, about $2 a night, inside weather's cap.
- Earliest effect: about Nov 6. Small dollars, but a second rule that works for a different reason.
3. The 10 AM gold dip goes live only through a fixed gate
- Why: the study found a real-looking dip at 10:00 AM, but it was one window out of 36 tested, so it needs a forward record.
- Gate, read about Nov 20: at least 20 weekdays; positive after fees in both halves; 10:00 AM beats the 11:00 AM variant in the week of Oct 26; average move at least 3 bp net. Then gold only, $2.50 risk, inside the engine's limits; back to paper after 6 R down from its best.
What has to be true before each step up
| Step | Must be true |
|---|---|
| Weather stage 1 → 2 | 30 settled live picks, net at least +$1.50, no glitch; the ladder code built; about $45 of predictions cash at 6 PM; the trader signs the stage-2 permission |
| Weather 2 → 3 → 4 | The same at each stage (+$3.00, then +$6.00), with about $90, then $180, of cash |
| Perps engine: more money or risk | At least 60 closed live-or-paper ideas whose average is above 0 R with 95% confidence |
| Rain to stage 1 | Change 2's gate |
| Miami hourly | Its Oct 30 look holds (+2¢ or better over 10+ dates), then 30 more dates; it shares weather's cap |
| Mentions long-shot No | Not before its 30-date check; it's losing today |
| Options | Hundreds of closes, and nothing before the review's re-check around Nov 6 |
Added after the big-move study
- Five more intraday setups fail after costs: VWAP reclaim or reject, squeeze breakout, hourly opening-range breakout, fade to VWAP, and a Bollinger fade. All 25 tests were negative in both the training and test periods; each is about even before costs, then loses 0.05–0.24 R a trade to fees, spread and slippage.
- Timing is real; direction isn't: big moves cluster 8–11 AM Eastern, busiest at 10 AM, and are several times likelier around scheduled releases. That supports watching the 10 AM gold dip, but it isn't evidence for it.
- No change to the allocation. The perps engine keeps its size; the order-book recorder on the paper engine needs about 30 ideas before it can inform sizing.
Tonight's edge hunt: leads and paper desks
Agents explore freely; a fixed referee scores each rule once on held-back data nobody has seen, and every look is recorded. Passing rules go to paper, then to real money through a written gate the trader signs.
| Line | Practice-data result | Where it stands | Gate to real money |
|---|---|---|---|
| Rain: government forecast vs price (the evening before) | 259 picks over 60 days, +6.95¢ a contract, both halves positive; last two weeks weak; its probabilities run ~5 points high | paper from Sat 6 PM; referee test after the forecast input is built | 60 picks / 15 days, ≥ +3¢, calibration within 10 points; at most $3 a night inside weather's limit |
| Gas prices: last-hour favourite (70–95¢, an hour before close) | 1,129 picks over 61 days, +4.82¢, about a 1-in-1,000 chance of luck; noon is flat, so it's a late-day effect | paper; referee test after the data rebuild | 120 picks / 10 days, ≥ +2¢; then 1 contract, at most $6 a night, its own $3 daily stop; earliest ~Nov 4 |
| Sports: combo seller (opposite 2–4 leg combo buyers) | Combo buyers pay ~3% (2 legs) to ~24% (10–19 legs) over the legs' worth | paper since Fri 8:59 PM | 2,000 picks over 15+ days at ≥ +1¢, then a practice run on real quote requests |
| Sports: NFL player props priced from Pinnacle (two model versions side by side) | Kalshi's ladder shape is closer to reality than a textbook curve, so version 1.1 uses Kalshi's shape with Pinnacle's centre | paper from Sun 3:29 AM | 3,000 strikes over 8 Sundays beating Kalshi's accuracy, plus 300 fills at ≥ +1¢ |
| Sports: resting bids anchored on Pinnacle (winners; spreads and totals incl. alternate lines) | Kalshi sits within ~0.5¢ of Pinnacle, so only the maker side can earn | switching on over the weekend | 400 fills over 20+ days at ≥ +0.5¢, beating Pinnacle's close |
| 15-minute commodities favourite | 8,609 picks, +0.64¢ at Kalshi's real fees, but ≈ 0 under the referee's fee math | waiting on a fee fix | — |
| Index ranges vs options; hourly temperatures; Bollinger shorts on stocks; intraday perps setups | All lost or no better than chance | dropped | — |
- New outside data now collecting: Google's WeatherNext 3 forecasts (access approved; usable from July 2026, the model's training cutoff), government blended forecasts back to 2021, a twice-daily forecast archive, first-print economic numbers, and a restarted Pinnacle and multi-book line recorder with props and totals.
- Confidence scoring: every pick now logs its win probability before the result. Sizing by confidence (a fraction of the Kelly bet, with caps) waits until about 100 settled picks show the probabilities beat the market's.
- Tonight's weather picks show why: by the agent's own numbers, about half of the eight live picks were worth less than their price.
The whole Kalshi landscape
A census of every market family from public data: 15,387 series, 170,437 open markets, 1.2 million October trades, and the referee's Aug 16 – Oct 16 quote store. "Taker" means crossing the spread and paying the fee; "maker" means resting a quote.
- Taking at random costs about 2–3¢ a contract in every family, mostly the fee. Makers collect most of what takers lose, unless the other side is a fast bot.
- Favourites pay only where the outcome is a public measurement that firms up before the close: weather and gas, at 70–95¢, +2 to +8¢ after fees. Long shots lose there by 4–25¢. In sports, props, mentions and crypto, the price is already right.
- Where a strong outside reference exists, Kalshi already matches it: Pinnacle on sports winners, and options on the daily S&P and Nasdaq ranges.
- Capacity is in sports: game winners trade about 50 million contracts a day. Weather and gas books hold about 10 contracts at the best price: enough for us at 1–8, not for big money.
| Rank | Family | Why it might work | Why not, yet | Status |
|---|---|---|---|---|
| 1 | Weather (highs and lows; rain, hourly, same-day next) | Favourites pay in both samples; the outcome is forecastable from public models, now including Google's WeatherNext 3 | Thin books; one month of live record | live, stage 1 and a rain-forecast lead (+7¢ a contract in practice data) waiting on the referee |
| 2 | Sports, maker side (resting bids anchored on Pinnacle; selling 2–4 leg combos) | The biggest books on Kalshi; takers lose about the fee; combos are overpriced against their legs | Fill risk: a resting bid gets hit when it's wrong | paper; spreads and totals added |
| 3 | Gas prices, last-hour favourite | +3 to +8¢ at 70–95¢ an hour before close | The evening-before version loses 5.6¢ a pick | being tested |
| 4 | Economic releases and the Fed | Public consensus and nowcasts; scheduled | A dozen events a year each | needs older data |
| 5 | Stock index ranges vs options | Options give a strong reference | Kalshi was sharper than options on every test day | dead |
| 6 | Culture with early public data | Some outcomes build up in public | Small, untested | untested |
| 7 | Mentions | — | Every price bucket at or below +0.5¢ after fees | losing |
| 8 | Sports props | — | No anchor, every bucket negative, wide spreads; an edge would have to come from a model | model research |
| 9 | Crypto and commodity 15-minute, crypto ranges | — | Fast bots; makers earned about zero | skip |
| 10 | Politics and long-dated futures | Large open interest | Money locked for months, no measured edge | skip |
The honest expected value
- Weather at stage 1, about 7 picks a night: roughly +$0 to +$0.35 a night. At stage 4 (8 contracts): +$0 to +$2.80 a night.
- Perps engine: roughly −$0.30 a day, give or take a few dollars.
- This is a plan to find out which lines deserve the money, at a cost of a few dollars. It is not an income yet.