HondaCivic

Wallet: 0x15ceffed7bf820cd2d90f90ea24ae9909f5cd5fa Window: 2026-04-19 to 2026-05-15 (27 calendar days, 27 active) Universe: 3,887 trades across 895 markets, 436 events · $705,395 BUY notional · $9,165 SELL notional Net P/L (resolved BUYs): +$7,596 on $705,364 deployed = +1.08% ROI in 27 days

The Punchline

HondaCivic is a weather prediction market specialist with a two-gear strategy: a high-volume penny carry trade that earns $0.001 per winning share on near-certain outcomes, layered with occasional high-conviction directional bets on specific temperature thresholds priced cheap by the market.

Every market in the book asks the same structural question: "Will the highest temperature in [City] be [X]°C/°F on [Date]?" He sweeps the entire probability distribution of that question for each city+date event, buying "No" at $0.999 on thresholds his weather model says won't be hit, and occasionally buying "Yes" at $0.08-$0.40 on thresholds he believes the market underprices. The 92.65% resolved win rate reflects the success of the "No" side strategy. The +1.08% headline ROI, while appearing modest, represents real economic return on a strategy designed to generate safe penny-level returns at scale.

The P/L is misleadingly compressed by the accounting method. The true economic picture is: 98.4% of capital is deployed at $0.999 earning a 0.1% carry per resolved position. On a mark-to-market basis, the strategy is approximately risk-free on those positions. The real variance and the meaningful alpha live in the 1.6% of capital allocated to sub-$0.50 "Yes" buys, where the April 30 London 18°C trade alone generated +$4,972.

The wallet is not a longshot bot, not a market maker (0.67% both-sides rate, and those 6 markets all had paired costs far above $1.00 indicating accidental rather than intentional pairing), not a copy-trader, and not a DCA accumulator. It is a weather forecasting arbitrageur who has built a model that tells him which temperature outcomes are near-certain and which are underpriced longshots, then executes systematically across a global city universe.

---

What He Trades

The complete trading universe is weather temperature markets on Polymarket. The pattern from the CSV is unambiguous:

"Will the highest temperature in [City] be [Temp] on [Date]?"
"Will the highest temperature in [City] be [Temp] or higher/lower on [Date]?"
"Will the highest temperature in [City] be between [Temp1]-[Temp2]°F on [Date]?"

Cities confirmed in the CSV sample: London, Moscow, Paris, Madrid, Istanbul, Warsaw, Amsterdam, Helsinki, Toronto, New York City, Chicago, Miami, Atlanta, Seattle, San Francisco, Los Angeles, Denver, Houston, Buenos Aires, Sao Paulo, Lagos, Wellington, Ankara, Milan, Munich (likely).

A single soccer bet on some non-weather market for $2.05 lost. Everything else is weather.

Market structure per city+date event: Each weather event spawns multiple markets, one per temperature threshold. For the "Buenos Aires on April 19" event, the CSV shows him active in markets for 18°C or below, 20°C, 22°C, 24°C, and 25°C simultaneously. He buys "No" on the thresholds that are far from the expected outcome, and occasionally buys "Yes" on the threshold closest to his point forecast.

Entry price anatomy: The sub-bucket analysis is critical here:

98.4% of capital is concentrated at the $0.90-$1.00 entry band. This is the defining feature of the strategy.

---

The Order of Operations: One Event, Market by Market

The cleanest single-event trace is London temperature on April 30, 2026. This one event generated +$4,972 P/L and is the dominant contributor to the wallet's 27-day total.

The London April 30 event spawned multiple temperature markets. From the data:

The April 30 London high actually reached exactly 18°C. The market had priced this at 7.9-16.2% probability. He held 9-12 fills on the "Yes" side. When it resolved at $1.00, those shares paid out at a 6× to 12× return on the $0.079-$0.162 entry price.

Walk-through of the April 30 London 18°C trade (from best_markets_by_pnl data):

• 12 total trades on this market, $12,545 total volume, 9 resolved, 9 wins
• Buys confirmed in CSV at $0.1262 (115 shares, $15.14), $0.1616 (202 shares, $34.08), $0.1100 (17.73 shares, $2.04), $0.1200 (52.5 shares, $6.58)
• These were "Yes" buys on a market priced at 8-16% probability
• London's actual April 30 high: 18°C exactly
• Resolution: Yes wins, every share pays $1.00
• Net P/L on just this market: +$4,972

The remaining $7,572 of volume on that market was almost certainly "No" buys on the flanking thresholds within the London April 30 event that also resolved correctly.

This is the core insight: the "Yes" buy at $0.10-$0.16 on a temperature that actually hits generates the majority of the wallet's total 27-day P/L. One correct point forecast worth $51 of capital beats 27 days of penny harvesting on $700K of near-certainty capital.

---

Why It Works: The Math

The strategy has two distinct EV calculations:

Strategy A: The Penny Carry

Entry price: $0.999
Resolution payout (win): $1.000
Gross per-share profit (win): $0.001
Gross per-share loss (lose): -$0.999

For Strategy A to be EV-positive:
  EV = p_win * $0.001 - (1 - p_win) * $0.999 > 0
  p_win > 0.999 / 1.000 = 99.9%

So: you must be >99.9% confident the outcome is "No"
to make the penny carry even nominally EV-positive.

His observed win rate on the $0.90-$1.00 band is 100% (3,540 wins, 3,540 trades). So the realized return of +$2,015 on $694,564 deployed = +0.29% over 27 days is consistent with a penny carry that wins essentially every time. The absolute loss on any single position is bounded at the cost of the position (rarely more than $1,000).

Strategy B: The Longshot Directional

Entry price: ~$0.10 (typical "Yes" buy on target threshold)
Resolution payout (win): $1.000
Gross per-share profit (win): $0.900

The April 30 London trade:
  Avg entry: ~$0.127 on ~400 shares
  Deployed: ~$51
  Payout: ~$400 x $1.00 = ~$400
  Net P/L: ~+$349 per fill cluster (9 fill clusters = ~$4,972 total)

For the "Yes" strategy to be EV-positive:
  EV = p_win * (1 - p) - (1 - p_win) * p > 0
  where p = entry price
  p_win must exceed p (i.e., true probability > market price)

The model must be pricing London April 30 18°C at 35%+ true probability while the market sits at 8-16%. That gap is the edge. It requires a genuinely superior weather model with better uncertainty quantification than the crowd-sourced Polymarket orderbook.

Combined P/L decomposition:

The $0.30-$0.40 band delivers the outsized ROI (+185%) because that's where the successful moderate-probability "Yes" bets land (the April 30 London trade counted at the $0.30-$0.40 band boundary if averaged across the fill range). The longshot $0.00-$0.20 band loses money in aggregate because not every "Yes" bet hits - the losers (Seoul May 1, London 16°C April 23) drag the band negative.

---

Phase 1: Trader Profile

Scale and Activity:

• 3,887 total trades (3,867 buys, 20 sells) over 27 days
• $705,395 BUY notional across 895 unique markets / 436 unique events
• Active all 27 days (100% active days)
• ~144 trades per active day (moderate cadence)
• 20 SELLs across 27 days (essentially never exits before resolution)

Trade Size Distribution:

The mean-to-median ratio of 9.2 is extremely high, indicating severe right skew. The P95 of exactly $999 is a hard cap the bot applies to most large fills ($999 = 999 shares at $1.00, or 1001 shares at $0.999). The $9,815 max fill is an outlier that came from a single large "No" position on a near-certain outcome. The Lorenz curve shows 50% of trades carry only 1.9% of capital, while the top 1% carry 14.8%.

Execution Signature:

• Median inter-fill gap: 91 seconds (semi-automated, not pure bot)
• P10: 0.0 seconds (same-second multi-fills occur)
• P90: 2,054 seconds (~34 minutes between fills at the 90th percentile)
• 28.8% of fills within 10 seconds
• This is a mixed automation signature: fast bursts within a single market (the $999 + small fills in rapid succession), interspersed with long pauses between markets as the operator identifies the next target

Trading Hours (UTC):

• Trades in every hour of the day (24/7 operation, though sparse overnight)
• Peak hour: 15:00 UTC (316 trades, +$5,487 P/L) - enormous outlier, almost certainly driven by the London April 30 resolution
• Secondary peaks: 19:00 UTC (339 trades), 16:00 UTC (321 trades), 12:00 UTC (259 trades)
• Lowest hours: 02:00 UTC (5 trades), 01:00 UTC (7 trades)
• The bot runs nearly 24/7 but is thin overnight (01:00-05:00 UTC)

No hard sleep window. Unlike SirMartingale's clean 23:00-02:00 gap, this wallet has scattered fills around the clock. It is either fully automated or operated by someone across multiple time zones.

---

Phase 2: Core Strategy Identification

Both-sides participation rate: 0.67% (6 of 895 markets). These 6 are accidents, not intentional pairing. The median paired cost of $1.236 (well above $1.00) confirms there is no spread-capture intent - anyone deliberately market-making would maintain paired cost below $1.00. These 6 markets were likely cases where he bought "Yes" on one threshold and then separately bought "No" on an adjacent threshold within the same event structure, and the system counted them as both-sides.

Classification: DIRECTIONAL BETTOR with a specialized WEATHER FORECASTING ARBITRAGE edge.

He is NOT:

• A market maker (paired cost $1.24 average, no spread capture intent)
• A crypto trader (zero crypto exposure)
• A sports bettor (one accidental soccer bet)
• A latency arbitrageur (weather markets don't have latency-exploitable price updates)
• A DCA accumulator in the conventional sense (though he does accumulate on single markets across hours)

He IS:

• A directional bettor with a weather forecasting model
• A carry trader who earns the $0.001 spread on high-confidence "No" outcomes
• An opportunistic longshot buyer when his model flags specific temperature thresholds as underpriced

---

Phase 3: Dominance Ratio Analysis

Six markets have both sides. All 6 fall in the "3.0x+" dominance bucket with a 100% dominant-side win rate. Mean paired cost of $1.26 means paired cost analysis is not applicable - there is no intentional spread. The dominant side in these 6 markets was the "No" side (high-probability outcome), which won all 6 times. This is consistent with random both-sides exposure from the event structure, not a deliberate strategy.

Dominance analysis conclusion: not applicable as a strategy-identification tool for this trader. The 0.67% both-sides rate is noise.

---

Phase 4: Entry Price Analysis

The price distribution is the most concentrated in our dataset. 98.5% of capital sits in the $0.90-$1.00 band, nearly all at exactly $0.999.

The sub-bucket concentration check reveals the singular insight: the $0.999 price point holds the majority of all trades. The bot is bidding the top of the near-certainty zone, absorbing the market's available liquidity at the highest possible "No" price before resolution.

The win-rate calibration in the $0.90-$1.00 band is exactly what you'd expect: 100% win rate (3,540 of 3,540) on outcomes priced at 99%+. The market is correctly pricing near-certainty, and he is capturing the spread.

The $0.30-$0.40 band is the anomaly: 66 trades, 39.4% win rate, +$5,053 P/L. Win rate of 39.4% against implied probability of 30-40% is essentially fair odds. The outperformance comes from the outsized payout when those bets hit. The London April 30 18°C "Yes" buys at $0.079-$0.162 are in the $0.00-$0.20 band and show 0% win rate in aggregate because more of them lost than won across the full 27 days - but the one that hit was London April 30, and it paid $4,972. The P/L on the band is negative (-$752 combined on $0.00-$0.20) but the April 30 market's P/L feeds into the $0.30-$0.40 band accounting because the market-level aggregation captures the full outcome.

---

Phase 5: Category and Vertical Breakdown

Single-category book. The "Other" category captures all weather markets since none of them match the standard keyword classifications (sports, crypto, politics, etc.).

The interesting breakdown is geographic. Based on the CSV and top markets data:

London is the highest-P/L geography by a large margin, driven entirely by the April 30 event. Ankara is the second-best geography due to multiple successful moderate-probability "Yes" bets. Most other cities are pure carry contributors.

---

Phase 6: Timing and Execution Analysis

Hourly P/L (UTC) - notable hours:

The 15:00 UTC spike is an artifact of a single day's large payout. The 06:00 UTC hour has the worst win rate (55.6%) and is the single loss hour in the book (-$125). This is early morning European time when same-day weather forecasts are freshest but also when he may be acting on less-certain information for cities reporting early-morning temperatures.

Day of week:

Thursday's extraordinary ROI (+5.94%) is driven entirely by the week containing April 30 (a Thursday). Friday is the only negative day-of-week aggregate, with -$151 driven by losses on markets where his forecasts missed. Tuesday's lower win rate (80%) compared to other days reflects more exposure to moderate-probability "Yes" bets on those days.

Accumulation pattern: Within a single market, he fires multiple small fills over a 30-minute to 3-hour window. The April 19 Chicago 46-47°F market shows him entering at 15:52 with 15+ separate fills ranging from $0.83 to $1,553 over 30 minutes. The large fill at 15:52:10 ($1,553) followed immediately by smaller fills ($16, $12, $11, $7, $6, $5) is the signature of a bot that fires a large "anchor" fill and then accumulates residual liquidity.

---

Phase 7: Filter Experiments

The price 0.30-0.70 filter is the most revealing finding: 72 trades carrying only $3,117 of capital generate +$5,301 in P/L (170% ROI). This is the moderate-probability directional bet layer of the strategy - the London April 30 "Yes" buys and the Ankara trades sit in this band. If you could run only this filter, you'd extract 70% of the total P/L on 0.44% of the capital.

The high-conviction filter returns +$338 on $11,348 (2.97% ROI) - it captures the both-sides markets which are accidental pairings, not intentional plays. Not useful.

The hour filter modestly improves ROI from 1.08% to 1.24% by excluding the 06:00, 08:00, 11:00, and 13:00 UTC hours where win rates are lowest. The absolute P/L impact is small (-$195).

---

Phase 8: Rolling Window Consistency

100% of 15-day rolling windows are green. One 7-day window touches slightly negative (-$0.91 on April 20), driven by a single position that failed on the opening day. The cumulative trajectory shows a dramatic jump in week 18 (the London April 30 event), then a steady but slower climb in weeks 19-20.

Weekly P/L:

W16 (Apr 19 only):        +$75
W17 (Apr 20-26):          +$927
W18 (Apr 27-May 3):     +$5,388  [London Apr 30 event here]
W19 (May 4-10):           +$984
W20 (May 11-15):          +$223
Cumulative:              +$7,597

The jump from W17 to W18 (+$4,461 incremental) is entirely the London 18°C April 30 trade. Without it, the total 27-day P/L would be approximately +$2,625 on $705K deployed - a 0.37% return that more accurately characterizes the baseline penny carry strategy.

---

Phase 9: P/L Decomposition

The P/L decomposition is simple: 3,582 winning positions pay out their full share count at $1.00 minus their cost. 284 losing positions pay $0.00 and the cost is fully lost. The key math is that 3,540 of those winners are the $0.999 penny carry (earning $0.001 each), while 42 winners are in the $0.30-$0.90 band earning much more per share.

The April 30 London 18°C market ($4,972 P/L) represents 65% of the total book P/L from a single event. This is massive concentration. Strip it out and the 27-day return drops to +0.37%.

---

Phase 10: Strategy Specification Summary

One-sentence summary: A weather forecasting arbitrage operation that buys near-certain "No" outcomes at $0.999 across global temperature markets to earn a penny carry, while selectively buying "Yes" at sub-$0.40 on temperature thresholds where a proprietary NWP model indicates the market underprices the probability of the outcome occurring.

Edge source:

1. Carry trade: Earn $0.001 per winning share on near-certain "No" outcomes. Requires 99.9%+ confidence in the forecast, achievable with modern NWP models on temperature ranges far from the expected daily high.
2. Forecast mispricing: Identify specific temperature thresholds (e.g., 18°C on a borderline spring day in London) where the ensemble model assigns 30-40% probability and the market prices it at 8-15%. Buy "Yes" at the market price.

What works: Near-certain "No" buys globally (stable carry). Correct point forecasts on borderline thresholds (windfall when right). London markets appear particularly well-calibrated for this operator.

What drags: "Yes" buys on thresholds that don't hit (e.g., Seoul 20°C+ May 1, -$113; London 16°C April 23, -$59). The sub-$0.20 directional bets collectively lose money across the sample period, offset by the one large London win.

What replicators must understand: The penny carry earns safe but tiny absolute returns ($2,015 on $694K = 0.29%). The alpha is in the directional calls. You cannot replicate this strategy without a credible NWP data feed and the ability to identify which temperature thresholds are systematically underpriced by Polymarket's crowd.

Wallet: 0x15ceffed7bf820cd2d90f90ea24ae9909f5cd5fa Window: 2026-04-19 → 2026-05-15 (27 active / 27 calendar days) Methodology: Cash-flow P/L = -buy_usdc + sell_usdc + remaining_share_payout. Resolved shares settle at $1 (win) / $0 (loss); open positions marked at last price.

Phase 1 - Trader Profile

Scale

Trade-size distribution (USDC per fill)

Inter-trade gap, same (market, outcome)

Phase 2 & 3 - Both-Sides Participation, Dominance Curve

• Both-sides rate: 0.67% (6 of 895 markets)
• Median paired cost: $1.2364
• Mean paired cost: $1.2605
• Paired cost % under $1.00: 0.0%
• Paired cost % under $0.97: 0.0%
• Median 2nd-side hedge lag: 29242s

Dominance buckets

Phase 4 - Entry-Price Analysis

Phase 5 - Category & Vertical Breakdown

Phase 6 - Timing & Execution

Net P/L by hour (UTC)

Phase 8 - Rolling Window Consistency

• Rolling 7-day windows green: 26 of 27 (96.3%)
• Rolling 7-day P/L range: -$1 → +$5,792
• Rolling 15-day windows green: 26 of 27 (96.3%)
• Rolling 15-day P/L range: -$1 → +$6,640

Weekly P/L

Phase 9 - P/L Decomposition

Phase 10 - Top Markets by Volume

Top 10 winners by P/L

Top 10 losers by P/L

Report generated 2026-05-17 00:03 UTC.

Wallet: 0x15ceffed7bf820cd2d90f90ea24ae9909f5cd5fa Window: 2026-04-19 to 2026-05-15 (27 days) Baseline: 3,866 resolved BUYs · 92.65% WR · $705,364 deployed · +$7,597 P/L · +1.08% ROI

---

The headline result

The price band filter is the most revealing filter in this analysis, but not for the reason you might expect. Applying the standard "$0.30-$0.70 sweet spot" filter to this wallet isolates 72 trades carrying $3,117 of capital that generate +$5,301 of P/L (170% ROI). This is the highest ROI of any subset of the book, but it covers only 0.44% of the total capital deployed.

The implications are:

1. The $0.30-$0.70 filter correctly identifies the directional-bet layer as the highest-ROI subset
2. But in absolute dollar terms, this layer generates only $5,301 - you'd need to run the penny carry simultaneously to generate meaningful absolute returns
3. The "exclude worst hours" filter modestly improves ROI without sacrificing much P/L
4. The dominance filter and category filter are both inapplicable or no-ops

The single most important filter finding: if you want to replicate only the high-ROI portion of this strategy, focus exclusively on the $0.30-$0.70 band and skip the $0.999 penny carry entirely. You get 70% of the P/L on 0.44% of the capital. The catch is you need a good enough forecast model to identify which of those bets win.

---

Filter results table

---

Filter-by-filter commentary

1. Price band filter ($0.30-$0.70) → MEANINGFUL LIFT (ROI)

This filter does something unusual: it cuts 99.56% of the capital and 98.1% of the trades, yet retains 69.8% of the P/L.

The 72 trades in the $0.30-$0.70 band are almost entirely directional "Yes" bets on temperature thresholds where the operator's model indicates the market is underpriced. The London April 30 18°C "Yes" buys (at $0.079-$0.162) technically fall in the $0.00-$0.20 band, but the moderate-probability bets in the $0.30-$0.40 band include the Ankara trades (+$291, +$153, +$104) and Buenos Aires (+$168) which collectively drive the +$5,301 total.

The 37.5% win rate against an implied probability of 30-40% looks like fair odds at face value. But the payout when these hit (3× to 10× depending on entry price) creates a positive-EV book if your forecast accuracy exceeds the market's implied probability by even a few percentage points.

Verdict for replicators: If you have a weather forecast model and want to run only the high-ROI component of this strategy, applying the $0.30-$0.70 filter to your universe selection focuses capital on the bets that matter. The penny carry ($0.999 buys) is low-ROI capital-intensive infrastructure; the directional bets are the alpha. Filter in this band and skip the rest.

The caveat: 72 trades in 27 days means roughly 3 qualifying opportunities per day. The strategy is capacity-constrained at the interesting ROI level. You cannot scale the 170% ROI by deploying more capital - you'd just have more $0.999 carry trades at 0.29% ROI.

2. High-conviction dominance filter → NOT APPLICABLE

Both-sides rate of 0.67% (6 markets) means there are only 46 trades in the qualifying set, and those 6 markets are accidental pairings, not intentional both-sides plays. The filter selects the dominant side of those 6 markets - all 6 of which resolved for the dominant side (the "No" side, naturally, since near-certain outcomes win). P/L: +$338 on $11,348 = 2.97% ROI. Worse than baseline.

This filter is structurally inapplicable. The strategy has no intentional both-sides component, so dominance ranking adds no signal. The filter is returning a random subset of the near-certainty "No" buys that happened to share a market with a tiny exploratory "Yes" bet.

3. Category filter → NO-OP

100% of trades are "Other" (weather). Single-category book. The category filter is identity-equivalent to the baseline. The one Soccer trade for $2.05 that lost is removed, returning +$2.05 in improvement - negligible.

There is no category diversification to optimize here. If you wanted to apply a category filter, the useful version would be a city/geography filter: exclude cities where your NWP model has historically underperformed (e.g., Wellington if you lack good southern hemisphere model data), and over-weight cities where you have strong local model data (e.g., European capitals with dense observation networks).

4. Hour filter (exclude worst 4 hours: 06:00, 08:00, 11:00, 13:00 UTC) → MODEST LIFT

Excluding the four worst-performing hours improves win rate from 92.65% to 96.4% and ROI from 1.08% to 1.24%. The absolute P/L drops slightly (-$195) because some winning carry trades are excluded along with the losers. Net effect: modest ROI improvement at the cost of slightly lower absolute P/L.

The 06:00 UTC hour is the key problem hour: 124 trades, 55.6% win rate, -$125 P/L. This is early European morning (07:00 London time), when same-day temperature forecasts are being revised as overnight observations come in. The model may be less reliable at this hour because the latest NWP model run (typically the 00Z or 06Z run, available from 03:00-06:00 UTC) hasn't propagated into his trading yet, but he's still trading on the previous run.

The 08:00, 11:00, and 13:00 UTC hours also show lower win rates (76.9%, 76.7%, 81.8% respectively). These overlap with European late morning when temperature forecasts for same-day resolution markets are most uncertain (the temperature is actively evolving and the final high hasn't been reached yet).

Verdict for replicators: Consider avoiding same-day markets during the 06:00-14:00 UTC window for European cities, where forecast uncertainty is highest and your model is working with stale NWP data. Trade into the morning runs (the 06Z GFS/ECMWF run becomes available around 09:00-11:00 UTC) rather than before it.

5. Combined filter → MODEST STACKING BENEFIT

Stacking the $0.30-$0.70 price filter with the worst-hour exclusion yields 11 trades, 45.5% WR, $3,019 capital, +$5,259 P/L, 174.1% ROI. The ROI is marginally higher than the price-band filter alone (+170%), but the trade count drops to 11 (too thin to draw conclusions over 27 days). The stacking provides only $42 of additional P/L over the price filter alone, suggesting the hour filter doesn't add meaningful signal in the directional-bet subset.

At 11 trades in 27 days, the combined filter is too restrictive for a production strategy. The individual price filter at 72 trades is more practical.

---

What filters would add genuine value here

The standard PR&R filter battery is designed for general-purpose directional bettors and market makers. For a weather-specialist like this wallet, the relevant filters don't exist in the standard kit:

None of these can be computed from the trade CSV alone. They require access to the same underlying meteorological data that drives the strategy.

---

Bottom line for replication

Three concrete filter recommendations:

1. DO apply the $0.30-$0.70 price filter if you want to run only the high-ROI directional component. You get 170% ROI on the interesting bets and avoid tying up capital in the 0.29% penny carry. But you need the forecast model to earn that ROI - without it, you're just randomly buying temperature longshots.

1. DO avoid the 06:00-08:00 UTC window for same-day European city markets. This is when NWP data is stalest and the forecast uncertainty is highest relative to the market price. The observed win rate drops to 55-77% in these hours vs 98%+ at other hours.

1. DO NOT apply the dominance filter or category filter. Both are no-ops or mildly destructive for this specific wallet structure. The strategy doesn't have a both-sides component to rank, and it's single-category by design.

The most powerful "filter" available is qualitative: identify which city+date events have a large gap between your model-implied probability and the Polymarket price, and size into those. The mechanical filters above are refinements at the margin.

Source wallet: 0x15ceffed7bf820cd2d90f90ea24ae9909f5cd5fa Strategy: Weather temperature market carry trade with directional forecast overlay Reference book: $705,395 BUY notional · +$7,596 net P/L · +1.08% ROI in 27 days (carry baseline: ~+0.37% ex-London April 30; directional overlay adds the bulk of meaningful alpha)

---

One-paragraph operator brief

Build a system that monitors Polymarket's temperature prediction markets for global cities, uses professional NWP model output (ECMWF/GFS ensemble) to compute your own probability distribution for each city's daily maximum temperature, then executes two parallel strategies: (1) buy "No" at $0.999 on any temperature threshold your model assigns less than 0.5% probability of occurring, earning the $0.001/share carry at resolution; and (2) buy "Yes" at market price on any temperature threshold your model assigns materially higher probability than the orderbook price (i.e., model says 30%, market says 10%). Cap the carry position at $1,000/market. Let directional "Yes" bets run up to $5,000/market when model conviction is high. Hold everything to settlement. Never sell early. Avoid trading European city markets before the 06Z NWP run has been processed (before 09:00 UTC).

---

1. Market selection

City universe used in reference book (inferred from CSV):

You do not need all cities simultaneously. Start with 5-8 cities where you have high confidence in your NWP data and observation quality. London, Paris, Madrid, and NYC are the highest-liquidity markets. Ankara showed strong alpha in the reference book (multiple +$100-$291 directional wins) despite lower volume.

---

2. Entry logic

The strategy has two entirely different entry modes that must be coded separately:

Mode A: The Penny Carry ("No" buys)

def should_enter_carry(market, city, date, threshold, direction):
    # Only for "No" side entries
    if direction != "No":
        return False, 0
    
    # Compute model-implied probability of this threshold being the high
    model_prob = get_model_prob(city, date, threshold)
    
    # Only enter carry if model assigns <0.5% probability to this outcome
    if model_prob > 0.005:
        return False, 0
    
    # Check timing: for European cities, wait for 06Z NWP run
    if city_region(city) == "Europe" and utc_hour(now()) < 9:
        return False, 0
    
    # Check market price: must be $0.999 or higher available liquidity
    ask_price = market.no_side.best_ask
    if ask_price > 0.999:
        return False, 0
    
    # Size: up to $1,000/market, sliced into sub-$1,000 fills
    clip = min(1000.0, available_capital() * 0.001)
    return True, clip

Mode B: The Directional Forecast ("Yes" buys)

def should_enter_directional(market, city, date, threshold, direction):
    # Only for "Yes" side entries
    if direction != "Yes":
        return False, 0
    
    # Compute model-implied probability
    model_prob = get_model_prob(city, date, threshold)
    market_prob = market.yes_side.mid_price
    
    # Only enter if model probability materially exceeds market price
    # Threshold: model must be 2× the market price or more
    if model_prob < market_prob * 2.0:
        return False, 0
    
    # Absolute minimum model confidence: 15%
    if model_prob < 0.15:
        return False, 0
    
    # Absolute maximum market price to enter: $0.70
    # (above $0.70, even correct calls don't pay enough to justify forecast error risk)
    if market_prob > 0.70:
        return False, 0
    
    # Size: proportional to the gap between model and market probability
    # Larger gap = larger bet. Cap at $5,000/market.
    gap = model_prob - market_prob
    clip = min(5000.0, gap * 50000)  # $500 per 1% gap, capped at $5K
    return True, clip

---

3. Exit logic (hold to settlement)

This strategy does not exit before resolution. The reference book has only 20 SELL transactions across 3,887 total trades (0.5%), and those are likely manual corrections, not a systematic exit strategy.

Rationale for hold-to-settlement:

• Carry trades: the $0.001/share profit only materializes at resolution. Selling early at $0.999 recovers capital but earns nothing. There is no secondary market for near-certain "No" positions at above-$0.999 prices.
• Directional trades: if your model says 35% and the market prices at 10%, and over the subsequent hours the market moves to $0.25, you could sell at 2.5× your entry for a partial profit. But the reference operator doesn't do this - he holds for the full 10× payout if right, zero if wrong.

def manage_position(position):
    # No active management. Hold to settlement.
    # Exception: if the market moves against the directional bet to >$0.80 "No",
    # your "Yes" position is effectively worthless (other side at 80%+).
    # Consider a soft stop at this threshold to recover small residual value.
    if position.direction == "Yes":
        no_prob = position.market.no_side.mid_price
        if no_prob > 0.80:
            # Optional: sell "Yes" shares for whatever partial value remains
            # Reference operator does not do this, but it reduces max-loss tail
            pass
    
    # Otherwise: let the market resolve
    return "hold"

Settlement mechanics:

• Winning position: all shares pay $1.00, net profit = shares × ($1.00 - entry_price)
• Losing position: all shares pay $0.00, net loss = -USDC_spent
• No gas required at resolution on Polygon (automatic settlement)

---

4. Sizing model

Mode A (penny carry): The sizing constraint is capital efficiency. Deploying $1,000 at $0.999 earns $1.00 at settlement. You need millions of shares to generate meaningful absolute returns. The reference book ran $694K/month earning $2,015 from the carry alone (0.29%). This is infrastructure capital: you accept the low ROI in exchange for near-certainty.

Mode B (directional): This is where you size for expected value. The reference book implies the following sizing ladder based on model-market gap:

The April 30 London 18°C trade was at 8-16% market price vs (estimated) 35-45% model probability. The reference operator deployed approximately $51 of capital on that single event and earned +$4,972. At more aggressive sizing (say, $500-$1,000 on a 30%+ gap), the payout would have been ~$50,000-$100,000 on the same forecast.

Capital allocation across the two modes:

The directional P/L is lumpy and event-driven. You will have weeks with zero directional wins and weeks with one win that exceeds the entire carry return. The April 30 London event is the canonical example.

---

5. The NWP model requirement (the load-bearing infrastructure)

Without a weather model, you cannot execute Mode B and can only run the penny carry at 0.29% monthly ROI. The model is the strategy.

Minimum viable stack:

The critical question for Mode B is: "What does the ensemble say about the probability of each specific temperature threshold?" This is not a deterministic point forecast. It is the full probability distribution of daily maximum temperature, discretized to 1°C/1°F buckets.

For London April 30 18°C example:

• ECMWF ENS 51 members might show: 10 members forecast 17°C, 18 members forecast 18°C, 14 members forecast 19°C, 9 members other
• Model-implied probability of 18°C: 18/51 = 35%
• Market price at time of entry: 8-16%
• Gap: 19-27 percentage points
• Decision: enter Mode B with gap-weighted size

def compute_temp_probability(city, date, threshold_celsius, nwp_ensemble):
    """
    Given 51 ensemble members, each providing a daily max temperature forecast,
    return the probability that the observed daily max == threshold (±0.5°C).
    """
    matching_members = sum(
        1 for member in nwp_ensemble 
        if abs(member.daily_max(city, date) - threshold_celsius) <= 0.5
    )
    return matching_members / len(nwp_ensemble)

---

6. Per-city and per-market inventory management

Each city+date event spawns multiple markets. The reference operator sweeps all of them. The correct approach for a replicator:

Step 1: Download the latest NWP ensemble for the city and date.

Step 2: Compute the full probability distribution across all temperature thresholds (e.g., for London April 30: P(13°C)=2%, P(14°C)=4%, P(15°C)=7%, P(16°C)=11%, P(17°C)=20%, P(18°C)=35%, P(19°C)=16%, P(20°C)=5%).

Step 3: For each threshold market on Polymarket:

• If model prob < 0.5% and market "No" price is $0.999: enter Mode A (carry)
• If model prob > 2× market "Yes" price AND model prob > 15%: enter Mode B (directional)
• Otherwise: skip

Step 4: Execute. The reference operator accumulates across multiple fills over 30-180 minutes within each market. This is consistent with: (a) liquidity being thin for large clips, requiring multiple fills to reach target size; or (b) a time-averaging strategy to reduce entry price volatility.

Per-event inventory cap:

• Mode A (carry): cap at $1,000 per individual threshold market, up to $10,000 across all thresholds in one city+date event
• Mode B (directional): cap at $5,000 per individual threshold market, maximum one directional bet per city+date event (bet on the threshold with the highest model-market gap)

---

7. Operational requirements

---

8. Hour scheduling

---

9. Risk profile

Drawdown characteristics:

• Mode A: virtually no drawdown per position (bounded by clip size, near-100% win rate)
• Mode B: high binary variance. A month with zero correct directional calls returns only the carry (+0.29%). A month with two correct calls returns +10-30% on the directional capital.

Maximum credible single-day loss: A large Mode B position on the wrong side of a surprise weather event. At $5,000 max/event cap, maximum single-event loss is $5,000. Over a 27-day period, expected losing directional trades at the reference win rate of ~37% (on the $0.30-$0.40 band) would produce roughly 2-3 loss events per month.

---

10. Diagnostic checklist

Run weekly:

---

What this playbook deliberately does NOT include

• No $0.999 carry at scale without a model. Buying "No" at $0.999 without a verified model that tells you the outcome is <0.5% probable is random betting with near-certainty outcomes at 0.001× payout. Don't run Mode A unless you have the NWP infrastructure to confirm the near-certainty claim.
• No multi-day forecasts beyond 48 hours. NWP model skill for daily maximum temperature degrades sharply beyond 2 days. Buying "No" at $0.999 for a market resolving 5 days out is not a carry trade; it is a risk that your 0.5% model probability is actually 5-15% real probability.
• No fixed city roster. Do not hardcode a city list and mechanically trade it. Markets for cities where your model or observation quality is poor should be excluded, even if Polymarket lists them. Lagos and Wellington are edge cases: good-quality daily max observation, but NWP skill varies.
• No same-second bursts on Mode B. The penny carry can be multi-filled in rapid succession (walking the orderbook depth). Mode B directional bets should be entered more carefully: one or two fills over 10-30 minutes to time-average the entry price in case the orderbook reprices.
• No selling before settlement on Mode A. The $0.001/share profit only materializes at resolution. Selling Mode A positions early at $0.999 earns zero.
• No leverage. The strategy works at 1:1 capital. Adding leverage to the carry component multiplies the tiny ROI into something that looks appealing on paper but adds liquidation risk on a strategy that earns 0.29%/month.
• No extrapolation from one event. The April 30 London 18°C trade generated +$4,972 and accounts for 65% of 27-day P/L. This is a single successful forecast, not a repeatable mechanical signal. Build the infrastructure to identify the next one; do not assume London April 30 is a recurring pattern.

The strategy is fundamentally a weather forecasting business that happens to express its edge through Polymarket. The playbook is only executable if you have or can build a better-than-market temperature forecast model. Without that, the penny carry alone earns 0.29%/month on deployed capital - real money at $700K scale, but not worth the engineering effort at smaller scales.