Michi1

Wallet: 0xa89518aca5a633a79ad1e9737209c9689f83faac Window: 2026-07-28 to 2026-08-26 (30 calendar days, 22 active) Universe: 1,071 trades, 79 markets, 42 events, $51,134 gross turnover Account P/L (verified): +$9,127.80 | Trading P/L: +$5,420 on $30,409 deployed = +17.4% ROI

The Punchline

This is a weather trader with a meteorological edge. Every market in the book is a real-world temperature or precipitation outcome: "Will the highest temperature in Shenzhen be 33 degrees C on August 7?", "Will the lowest temperature in Seoul be 24 degrees C on August 23?", "Will Super Typhoon Dolphin hit China?". The strategy is not predicting sports results, political outcomes, or crypto prices. It is reading weather forecast data for Chinese and Korean cities and finding markets where Polymarket's crowd-implied probability diverges from what the forecast actually says.

The edge has two stacked components. The primary alpha mechanism is longshot accumulation at $0.001-$0.10: buying outcomes the market prices as near-impossible, holding to resolution, and collecting 10x-1000x payouts when the forecast proves correct. The secondary mechanism is near-risk-free spread capture via both-sides paired bets: buying Yes at $0.98-$0.999 and No at $0.001-$0.01 simultaneously, locking $0.009-$0.07 of guaranteed profit per share pair on markets where the outcome is already known or highly certain.

The account P/L of +$9,127.80 is Polymarket's verified total over 30 days. Trading P/L is +$5,420. The $3,638 gap is unexplained by measured incentives and is not described here as a reward. Week 34 (Aug 17-23) alone produced $5,516 of the $7,339 trading P/L, driven by two massive longshot hits: Seoul 27C (+$3,757) and Shenzhen low 28C (+$1,369). The strategy's P/L is lumpy and event-driven, not smooth.

---

What He Trades

The universe is weather markets for East Asian cities, almost exclusively, with a small satellite position in typhoon track markets during the active season.

City temperature markets (primary):

• Shenzhen high temperature: dozens of markets across the window, each asking "Will the highest temperature be exactly X degrees C on date Y?"
• Shenzhen low temperature: same structure
• Seoul (Incheon) high and low temperature
• Shanghai high and low temperature
• Hong Kong high temperature
• Jinan high temperature

Typhoon track markets (secondary, seasonal):

• Super Typhoon Dolphin: landfall target (China vs Japan), intensity at landfall
• Tropical Storm Saudel: landfall target, intensity

Structure of temperature markets: Each day's temperature event spawns multiple markets, one per candidate temperature value. For a given day in Shenzhen, there might be markets for "Will the high be 29C?", "30C?", "31C?", "32C?", "33C?", "34C?", "35C?", "36C?" all running simultaneously under the same event slug. Exactly one of these can resolve Yes. The others resolve No.

This structure creates the opportunity for both strategies: (a) buy the near-certain winner at $0.98+ for near-risk-free spread capture, and (b) buy the longshot candidates at $0.001-$0.10 in case the market has mislabeled which temperature will actually be hit.

---

The Order of Operations -- One Market, Trade by Trade

"Will the highest temperature in Seoul (Incheon) be 27 degrees C on August 21?" -- the best single market in the book.

This is the highest-P/L market: 10 trades, $274.15 deployed, +$3,756.80 P/L, 10/10 resolved wins. Michi1 bought Yes at multiple price points while the market priced 27C as near-impossible. The temperature hit exactly 27C.

Walk-through:

1. 08:10:58 -- Entry probe. First buy at $0.007 (700 basis points), 100 shares for $0.70. The market prices 27C as a 0.7% probability. Michi1's forecast data apparently says this is far too cheap.

1. 08:11:30 to 08:11:46 -- Small accumulation. A series of small buys at $0.01 totaling ~$2.13. These are exploratory -- walking the book to see how much liquidity exists at the floor.

1. 08:11:48 -- The conviction fill. Four buys fire in the same second, capped by a 1,646-share buy at $0.1277 for $219.37. This is the tell: Michi1 is not guessing at floor prices. He has enough confidence to put $219 into a market priced at 12.77% probability. Total average entry across all 10 trades: $0.1582 per share.

1. Resolution. Seoul's high temperature on August 21 was 27C. All 1,916+ shares pay $1.00. Gross proceeds: approximately $4,031. Net: +$3,757 on $274 deployed = +1,371% return.

The structure -- tiny probes at $0.01 followed by a large conviction fill at a slightly higher price -- appears across multiple other markets. It is the signature of a trader testing liquidity before committing the real size.

---

Why It Works -- The Math

The strategy has two positive-EV mechanisms operating simultaneously.

Mechanism 1: Spread capture on near-certain outcomes

Paired bet structure (both-sides markets):
  Buy Yes at $0.985 (near-certain winner)  = cost $0.985 per share
  Buy No  at $0.005 (near-certain loser)   = cost $0.005 per share
  Total paired cost                        = $0.990 per pair
  Payout on resolution (Yes wins)          = $1.000 (Yes) + $0.000 (No)
  Guaranteed profit per pair               = $0.010 per share (1.0% ROI)
  Observed mean paired cost: $0.9902
  Observed profit margin:    ~$0.0098 per pair

This mechanism requires knowing which temperature will win before the market resolves. The locked spread is tiny but risk-free when the forecast is correct. Across 18 paired markets, the dominant side won 18/18 times.

Mechanism 2: Longshot lottery with positive-EV entry

Representative longshot trade:
  Market: Seoul 27C on Aug 21
  Market implied probability: ~1.0-12.77% (entry prices ranged $0.007-$0.1277)
  True probability (per Michi1's forecast): presumably ~20-30%
  
  Entry: 1,916 shares at avg $0.143 = $274 deployed
  Payout on win (confirmed): 1,916 shares x $1.00 = $1,916+ gross
  Net P/L: +$3,757 (net after $274 deployed and including all fills)
  
Sub-$0.10 band aggregate (454 trades):
  Capital: $753
  P/L:     +$1,629
  ROI:     +216% 
  Win rate: 1.5% (6-7 wins out of 454 trades)
  Avg payout on a win: $1,629 / 7 wins = ~$233 per winning trade cluster

The EV math is straightforward: if the market prices an outcome at 1% and the true probability is 5%, the expected value of buying at $0.01 is $0.04 (a 4x return). Even at a 1.5% realized win rate, the 100x+ payouts dominate the economics. The concentration of 67% of all trades in the sub-$0.10 band confirms this is the primary volume mechanism.

---

Phase 1 -- Trader Profile

Scale and Activity

Trade size distribution: extreme power-law concentration

The median trade is $0.81 -- a single penny-level lottery ticket. The mean is $47.74 because a small number of large conviction positions dominate. The top 5% of trades carry 76.1% of all capital deployed. This is an extremely concentrated power-law book. The Lorenz curve confirms near-maximal inequality: the bottom 50% of trades carry only 0.15% of capital, while the top 1% carry 37.6%.

This structure is intentional. The tiny trades ($0.01-$1.00) are the longshot lottery tickets. The large trades ($200-$4,000) are the conviction positions where Michi1 is confident about the temperature outcome.

Execution Signature

The execution signature is semi-automated with burst patterns. The 39.7% sub-10-second fill rate and frequent same-second multi-fills (clearly visible in the Seoul 27C trace: four fills at 08:11:48) suggest either a bot or a rapid manual interface. But the 3,217-second mean (vs 18-second median) shows long inactive gaps between trading sessions, consistent with a human who monitors weather forecasts at specific times and then executes a burst of trades. This is not a 24/7 automated bot.

Trading Hours

Hour 08:00 UTC is by far the dominant hour: 237 trades and +$3,818 of P/L. This is 4:00pm China Standard Time (CST = UTC+8) and 5:00pm Korea Standard Time. Afternoon local time is when numerical weather prediction (NWP) model runs for the next day are typically published and when temperature forecasts for cities are most accurate. The trading schedule follows the forecast publication cycle, not a crypto session or a US business day.

Hours 00:00-03:00 UTC show zero to negligible P/L (0.0% win rate), confirming these are not active trading hours.

Archetype: WEATHER FORECASTER with dual mechanisms: paired-spread capture on near-certain outcomes and longshot accumulation on mispriced probabilities.

---

Phase 2 -- Core Strategy Identification

Both-sides participation: 22.8% (18 of 79 markets). This is substantial but not the dominant structure. The primary mode is directional (single-side bets on specific temperature outcomes).

Classification: The strategy blends two archetypes:

• B (Directional Betting) -- primary: 77% of markets are one-sided directional bets on specific temperature outcomes. These are not random directional calls -- they are informed by meteorological data that Michi1 believes is more accurate than the market's crowd pricing.

• A (Both-sides Spread Capture) -- secondary: 18 markets with both sides purchased at 3.0x+ dominance ratios. All 18 dominant sides won. The structure is deliberately asymmetric: large position on the near-certain outcome, tiny position on the loser.

What this is NOT:

• Not a crypto trader (zero crypto markets)
• Not a sports bettor
• Not a political markets trader
• Not a copy-trader (the markets are too niche for copy-following to make sense)
• Not a market maker (only 22.8% both-sides participation, and those are highly asymmetric rather than balanced)

The core edge is meteorological. Michi1 has access to weather forecast data (likely numerical weather prediction model outputs, meteorological APIs, or professional forecasting services) for Chinese and Korean cities and uses that data to identify markets where Polymarket's crowd-implied probability differs materially from the forecast probability.

---

Phase 3 -- Dominance Ratio Analysis

All 18 both-sides markets fall into the 3.0x+ bucket. There are zero markets in the 1.0-1.5x, 1.5-2.0x, or 2.0-3.0x ranges. This is structurally significant: Michi1 never pairs bets unless he is extremely confident in the direction. He does not hedge with balanced positions or modest tilts. When he pairs, he pairs hard.

The 100% dominant-side win rate on 18 markets is the most important single statistic in this phase. It confirms that the pairing behavior is not market-making (which would show ~50% win rates on the dominant side) but rather certainty-based directional betting with a small hedge. When Michi1 buys both sides at 3x+ dominance, he already knows (or is extremely confident) which side will win. The tiny non-dominant purchase ($0.001-$0.01 per share) is either a hedge against surprise or a mechanism to participate in the spread capture math.

The mean paired cost of $0.9902 (1.0% below $1.00) is the guaranteed profit per pair. Across the 18 markets, the spread income was $80.28 -- modest in absolute terms but a true zero-risk profit stream.

---

Phase 4 -- Entry Price Analysis

The ROI inversion is the central finding of this phase. The $0.90-$1.00 band holds 71% of all BUY capital ($21,707 of $30,409) but generates only $550 of P/L -- a 2.5% ROI. These are the near-certain winner purchases in the paired-spread markets plus any other high-probability directional bets.

Meanwhile, the sub-$0.20 bands hold only 4% of capital ($1,261 combined) but generate +$4,905 of P/L -- 90% of total trading profit on 4% of capital. The $0.10-$0.20 band is the standout: 16 trades, $508 deployed, +$3,276 P/L, +645% ROI. This band includes the bulk of the Seoul 27C position (entry prices of $0.10-$0.13).

Sub-bucket inspection: The per-cent histogram reveals a striking bimodal distribution. The single largest concentration of trades is at $0.001 (hundreds of penny-level buys) and a second cluster at $0.98-$0.999 (the near-certain winner purchases). The coin-flip zone ($0.40-$0.60) has almost zero activity (only 2 trades). This trader never bets at fair odds. He either knows the answer (and buys near-certainty) or thinks the market is dramatically wrong (and buys the longshot).

---

Phase 5 -- Category and Market-Type Breakdown

The entire book is categorized as "Other" by the standard framework (no weather category). The meaningful breakdown is by city and market type:

Top 5 markets by P/L:

The P/L is extremely concentrated. The top 2 markets account for +$5,126 of the $7,339 total trading P/L -- 70% of all profit from 2% of total market count.

Typhoon markets as a secondary vertical: The Typhoon Dolphin and Tropical Storm Saudel positions show Michi1 extending his meteorological edge from temperature forecasting to tropical cyclone tracking. The Dolphin markets returned +$760 (intensity at Japan landfall), +$472 (China landfall), and +$265 (Japan landfall No side). This is the same underlying skill -- reading numerical weather models -- applied to a different atmospheric phenomenon.

---

Phase 6 -- Timing and Execution Analysis

Peak hours (UTC) by P/L:

Weakest hours: Hours 00:00-03:00 UTC: zero wins, negative or near-zero P/L across all four hours combined (-$0.81 total).

The 08:00 UTC peak is 4pm CST / 5pm KST -- afternoon local time in the target cities. Day-ahead weather forecasts in China and Korea are typically updated at 00:00 UTC (8am CST model run) and 12:00 UTC (8pm CST model run), with forecasters publishing human-readable summaries in the afternoon. The 08:00 UTC burst likely corresponds to Michi1 receiving updated forecast data and executing positions.

Day-of-week P/L:

Friday's extraordinary +156.7% ROI is almost entirely from the August 21 Seoul 27C win (which fell on a Friday). The day-of-week pattern is dominated by event timing rather than any structural weekday vs. weekend dynamic.

Accumulation window: The typical burst pattern within a single market spans 10-30 minutes. The Seoul 27C example shows all 10 trades compressed into 10 minutes and 18 seconds (08:10:58 to 08:11:48 UTC). This is consistent with a human executing rapidly once the forecast data arrives, not a slow DCA accumulation.

Second-side lag: The median lag between first and second side in both-sides markets is 2.0 seconds. Near-simultaneous entry -- both the near-certain winner and the longshot loser are purchased within 2 seconds of each other. This is the paired-spread execution signature: one decision, two near-simultaneous fills.

---

Phase 7 -- Filter Experiments

The standard filter battery is almost entirely destructive here.

The price $0.30-$0.70 filter reduces 653 trades to 8 and P/L from $7,339 to $478 -- a 93% reduction in profit. This filter eliminates the entire longshot engine (sub-$0.10 and $0.10-$0.20 bands where 90% of P/L is generated) and also eliminates the near-certain winner band ($0.90-$1.00). The only trades it keeps are the handful of mid-probability directional calls, which happen to win often (87.5% win rate on 8 trades) but generate negligible absolute P/L.

The high-conviction filter (dominant side only, dominance 2x+) reduces P/L to $428 at a meager 3.6% ROI. These 114 trades are almost all the near-certain winner purchases at $0.98-$0.999 -- they win 100% of the time but pay only 0.2-2.5% gross return.

The hour exclusion filter is effectively a no-op: excluding hours 00-03 UTC removes 10 trades worth $78 of capital and $4 of P/L. Michi1 barely trades in those hours anyway.

The single useful dimension for filtering is within the longshot zone: focusing on the $0.10-$0.20 band (16 trades, +645% ROI) over the sub-$0.10 band (454 trades, +216% ROI). The $0.10-$0.20 band has higher per-dollar ROI, fewer trades to manage, and is where the Seoul 27C conviction position was placed. A replicator who could only access the $0.10-$0.20 band would capture the highest-density alpha.

---

Phase 8 -- Rolling Window Consistency

100% of 15-day rolling windows are green. The cumulative P/L line is positive throughout and accelerates sharply in week 34. However, 86.7% of 7-day windows are green -- four 7-day windows are negative, corresponding to the mid-August period (Aug 17-20) where there were no resolved wins and small losses on temperature bets.

The weekly stair pattern is highly non-linear:

Week 33 (Aug 12-14) had a 2.1% win rate -- nearly all losses -- and still managed +$95 of P/L. This is the spread-capture mechanism absorbing the directional miss: even when the longshots don't hit, the near-certain-winner purchases prevent the week from going deeply negative.

---

Phase 9 -- P/L Decomposition

The hedge tax of -$170.26 (cost of the tiny losing-side positions in the 18 paired markets) exceeds the gross spread income of $80.28 by $89.98. This means the spread-capture mechanism is slightly net-negative in isolation. The paired bets only make economic sense as a combined structure: the near-certain winner purchase generates the real return (settling at $1.00), while both the spread income and the hedge tax are secondary effects.

The $3,638 unexplained gap represents positions that are open at the time of measurement (mark-to-market) or basis differences in how Polymarket attributes series P/L. It is not attributable to any identified reward program and is not described here as such.

The most important decomposition finding: the longshot directional bets (primarily sub-$0.20 entries) are responsible for essentially all of the trading alpha. The spread mechanism produces negligible net income ($80 minus $170 hedge tax = -$90). The near-certain winner purchases return 2.5% on 71% of capital. The entire +$5,420 of meaningful P/L flows from the cases where Michi1's forecast was correct and the market had dramatically mispriced the probability.

---

Phase 10 -- Strategy Specification

One-sentence summary: A meteorologically-informed Polymarket weather trader who accumulates longshot positions on temperature outcomes in East Asian cities when numerical weather models show higher-than-market probability, while simultaneously running a paired-spread capture on near-certain outcomes.

Edge source: Real-time or near-real-time numerical weather prediction (NWP) model data for Chinese and Korean cities, used to identify markets where Polymarket crowd pricing diverges from the forecast probability by more than 5x.

Market selection: Daily temperature markets (high and low) for Shenzhen, Seoul, Hong Kong, Shanghai, and Jinan. Secondary: tropical cyclone track and intensity markets during typhoon season (June-November).

Entry logic: When forecast probability for outcome X is materially higher than market price: buy at market price (typically $0.01-$0.20). When outcome X is near-certain per forecast: buy near-certainty side at $0.98-$0.999 and also buy the loser at $0.001-$0.01 for paired-spread capture.

Sizing model: Power-law. Most bets are $0.01-$1.00 penny tickets. Conviction positions are $100-$4,000 when the forecast-to-market gap is large. Top 5% of trades by size carry 76% of total capital.

Exit strategy: Primarily hold to resolution. Some active SELL activity (397 sells vs 674 buys) suggests partial profit-taking or position management when price moves significantly before resolution.

Risk management: The paired-spread mechanism provides a floor: even on bad directional weeks, the near-certain winner purchases prevent deep drawdowns. Worst single week was flat-to-positive. Worst observed rolling 7-day window was -$103.

Replication parameters: Detailed in the playbook tab.

Wallet: 0xa89518aca5a633a79ad1e9737209c9689f83faac Window: 2026-07-28 → 2026-08-26 (22 active / 30 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: 22.78% (18 of 79 markets)
• Median paired cost: $0.9951
• Mean paired cost: $0.9902
• Paired cost % under $1.00: 94.4%
• Paired cost % under $0.97: 5.6%
• Median 2nd-side hedge lag: 2s

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 30 (86.7%)
• Rolling 7-day P/L range: -$103 → +$5,640
• Rolling 15-day windows green: 30 of 30 (100.0%)
• Rolling 15-day P/L range: +$10 → +$5,611

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 2026-08-27 17:50 UTC.

Wallet: 0xa89518aca5a633a79ad1e9737209c9689f83faac Window: 2026-07-28 to 2026-08-26 Baseline: 653 resolved BUYs, 28.6% WR, $30,409 deployed, +$7,339 P/L, +24.1% ROI

---

The headline result

Every standard filter destroys more than 85% of the profit. The price filter, the high-conviction filter, the category filter, and the hour filter are all either no-ops or deeply harmful when applied to this book. The reason is structural: Michi1's alpha lives in the sub-$0.20 entry zone (longshot directional bets on meteorological mispricing) and in the near-$1.00 zone (spread capture on near-certain outcomes). The canonical filter sweet spots ($0.30-$0.70, dominance 2x+) capture neither of these zones.

The most important filter insight for a replicator is negative: do not filter this strategy at all using the standard battery. The three genuinely useful analytical refinements come from within-zone analysis, not from the standard filter set.

---

Filter results table

---

Filter-by-filter commentary

1. Price band filter ($0.30-$0.70) -- DESTRUCTIVE

This is the most harmful filter in the battery applied to this trader. Applying the $0.30-$0.70 sweet-spot range reduces 653 trades to 8 and P/L from +$7,339 to +$478 -- a 93.5% reduction in profit. ROI also deteriorates: 7.8% vs 24.1% at baseline.

The mechanism of destruction: the filter eliminates the entire sub-$0.20 entry zone (454 + 16 = 470 trades, holding +$4,905 of P/L, 67% of all profit) AND the entire $0.90+ zone (143 trades, $21,707 of capital, +$550 P/L). What remains is the 8-trade coin-flip zone, which wins 87.5% of the time but generates trivial absolute P/L.

The reason this filter works on most books but fails here: standard directional betters on Polymarket make money by correctly calling coin-flip outcomes better than the crowd. Michi1 makes money by identifying dramatic probability mispricings at the extremes. The sweet-spot filter is built for the former case and is actively hostile to the latter.

2. High-conviction filter (dom 2x+, dominant side only) -- DESTRUCTIVE

The high-conviction filter keeps 114 trades -- all of them the near-certain winner purchases at $0.98-$0.999 in the 18 paired markets. These win 100% of the time (by construction: they are the dominant side in markets where Michi1 already knows the answer) but generate only +$428 at 3.6% ROI. This is a 94% reduction in P/L vs baseline.

The filter fails here because the word "high-conviction" is operationally misleading for this trader. The highest-conviction trades are the 10 Seoul 27C buys at $0.007-$0.1277 -- but those are on the non-dominant side of the paired structure (the market priced them as near-impossible) and fall in the sub-$0.20 band, not the 3x+ dominance band. The filter catches the wrong thing.

The 100% dominant-side win rate (18/18) tells a more useful story about the quality of this trader's meteorological knowledge than it does about a filterable edge -- it is evidence that the strategy works, not a filter criterion.

3. Category filter (top category: Other) -- NO-OP

100% of trades are in the "Other" category. The filter is identity-equivalent to baseline. P/L delta: $0.

This is a structural limitation of the framework, not a finding about the strategy. The appropriate category breakdown for this trader is by city and market type (Shenzhen temp, Seoul temp, typhoon track), not by the standard category labels.

4. Hour exclusion filter (exclude hours 0-3 UTC) -- NO-OP

Excluding the four worst-performing hours (00:00-03:00 UTC) removes 10 trades, $78 of capital, and $4 of P/L. The delta vs baseline is negligible: -$4 on $7,339 (0.05%). Michi1 already barely trades in these hours. The filter has nothing to remove that isn't already near-zero.

The hourly P/L pattern confirms the operator is active during afternoon CST (08:00 UTC = 4pm CST) and mostly inactive from midnight to 3am CST (16:00-19:00 UTC). Forcing a schedule change via filtering would not improve anything -- the operator's natural schedule already avoids the low-edge hours.

5. Combined filter -- DESTRUCTIVE

The combined filter (price $0.30-$0.70 + high-conviction) is dominated by the price filter, returning the same 8 trades and +$478 P/L as the price filter alone. Stacking filters here doesn't add refinement -- it just confirms that the price filter is the binding constraint and it destroys the book.

---

What filters would actually help

The standard battery is designed for crypto betters and sports bettors where the coin-flip zone is the alpha zone. For a weather trader operating at price extremes, none of the standard dimensions apply. Three genuinely useful refinements are available but require data beyond the standard filter set:

---

Bottom line

The standard PR&R filter battery is not aligned with this strategy's structure. Three concrete take-aways:

1. DO NOT apply any price band filter. Both the sub-$0.20 zone (where the real alpha lives) and the $0.90+ zone (spread capture) are outside the standard sweet spot. Filtering either destroys the strategy.

1. DO NOT apply the high-conviction dominance filter. The 100% win rate on dominant-side picks is impressive evidence that the strategy works -- it is not a filter to apply mechanically. Keeping only the dominant-side picks removes the longshot engine that generates 90% of trading P/L.

1. DO consider within-zone prioritization. Focus capital allocation on the $0.10-$0.20 entry tier when a strong meteorological signal is present, rather than spreading penny bets across the full price spectrum. This is the highest-ROI refinement available from the data.

The playbook tab has the full implementation specification including how to size across the two price zones.

Source wallet: 0xa89518aca5a633a79ad1e9737209c9689f83faac Strategy: Meteorologically-informed weather market trading on Polymarket, with dual mechanisms: longshot accumulation on temperature mispricings and near-risk-free spread capture on near-certain outcomes. Reference book: $30,409 BUY notional deployed, +$5,420 trading P/L (+17.4% ROI on resolved trades), +$9,127.80 account total (verified by Polymarket, 30 days).

---

One-paragraph operator brief

Build or run a Polymarket weather trading operation focused on daily temperature markets for East Asian cities (Shenzhen, Seoul, Hong Kong, Shanghai) and typhoon track markets during the active season (June-November). The core signal is day-ahead numerical weather prediction (NWP) model output for the target cities, compared against Polymarket's crowd-implied probability for each candidate temperature. When your forecast says outcome X has a 15-25% probability and the market prices it at 1-3%, buy that longshot at $0.01-$0.20. Separately, when your forecast says outcome Y is 95-99% likely, buy Y at $0.97-$0.999 AND buy the competing outcomes at $0.001-$0.01 to lock a paired spread. Size your longshot buys proportionally to your forecast confidence: penny bets when uncertain, $100-$500 when the model is strongly divergent from market pricing. Expect monthly ROI of 15-25% on deployed capital, with occasional outsized weeks when a major mispriced event resolves in your favor.

---

1. Market selection

City prioritization: Seoul appears to carry the widest mispricings based on the reference book (one Seoul market returned +$3,757 on $274 deployed). Shenzhen is the workhorse for volume and spread capture but has tighter pricing. The priority order for capital allocation is: Seoul high-divergence signals first, then Shenzhen confirmed signals, then Shanghai and Hong Kong.

Market structure note: For a given city and date, Polymarket typically lists 6-12 markets covering each candidate temperature value (e.g., "Will the high be 27C?", "28C?", "29C?", ... "36C?"). These markets are mutually exclusive within the event -- exactly one resolves Yes. This structure is fundamental to the strategy because it enables both mechanisms simultaneously.

---

2. Entry logic

def should_enter_longshot(market, forecast_prob, market_price):
    # Divergence threshold: forecast must be at least 5x the market price
    if forecast_prob < market_price * 5:
        return False  # not enough mispricing
    
    # Price zone: only buy longshots below $0.25
    if market_price > 0.25:
        return False  # too expensive for longshot approach
    
    # Market must resolve within 5 days
    if days_until_resolution(market) > 5:
        return False
    
    # Don't buy if market liquidity is insufficient to absorb your clip
    if market.available_liquidity < target_clip * 0.5:
        return False
    
    return True

def should_enter_spread_capture(market, event, forecast_prob):
    # Near-certain winner: forecast says >90% probability
    if forecast_prob < 0.90:
        return None
    
    # Buy the near-certain outcome at whatever the book offers near $0.97+
    # Simultaneously buy the competing outcomes at $0.001-$0.01 each
    # Target combined paired cost below $1.00
    return "pair"

Entry timing: The reference book peaks at 08:00 UTC (4pm China Standard Time). Day-ahead forecast data for Chinese and Korean cities is most complete in the afternoon local time. Enter after you have the updated NWP model output for tomorrow's temperatures, typically by 15:00-18:00 local CST (07:00-10:00 UTC). Do not enter during the 22:00-06:00 UTC period when forecast data is stale and market activity is low.

Multi-fill execution: The reference trader fans out across the orderbook with multiple sequential fills within 30-60 seconds. Small fills at floor price ($0.001-$0.01) to confirm liquidity, then a larger conviction fill at a slightly higher price if the book has depth. The Seoul 27C example shows this exactly: 9 small fills ($0.007-$0.02) followed by one large fill (1,646 shares at $0.1277 for $219).

---

3. Exit logic

The primary exit is hold to resolution. Temperature markets resolve at a specific time (typically the day after market creation, when the official temperature reading is published). You hold until resolution.

Active exit conditions (SELL before resolution):

Trigger 1: Price has moved dramatically in your favor (longshot priced at $0.001 
            now at $0.30+) AND market still has 24+ hours to resolution.
            Action: Sell 30-50% of position to lock realized gains. Hold rest.

Trigger 2: New forecast data arrives that contradicts your original thesis.
            Action: Exit the longshot position entirely if new probability < 3%.
            Hold the near-certain winner position regardless.

Trigger 3: Near-certain winner position -- if you can sell at $0.999 shortly
            before resolution with negligible loss, do so to avoid any
            late-stage counterparty risk.
            Action: Sell at $0.999 when available.

The reference book shows 397 SELLs against 674 BUYs -- a substantial amount of active position management. Many of these SELLs are at $0.999 (extracting near-full value on the near-certain winner positions) or at penny prices (tidying up small loser positions before they resolve at $0.00).

The core rule: do not exit a longshot position early unless the thesis has changed. The Saudel market example in the CSV shows complex in-and-out activity as the storm track evolved -- that is correct behavior when new meteorological data is continuously updating the forecast.

---

4. Sizing model

The reference book uses an extreme power-law structure. Replicate it:

Capacity ceiling: Individual temperature markets on Polymarket have limited liquidity. The reference trader's single largest position was $3,976. Above $5,000 per market, you will start moving prices. The natural capacity ceiling for this strategy is approximately $50,000-$100,000 of total bankroll, beyond which you would need to fragment across multiple wallets or accept degraded entry prices.

The sizing rule that matters most: The highest-ROI positions are the $100-$500 conviction longshots at $0.10-$0.20 entry price. The Seoul 27C position (avg $0.158 entry, $274 deployed) returned +$3,757. These are not the most frequent trade type by count (only 16 trades in the $0.10-$0.20 band vs 454 in sub-$0.10), but they generate 3x the ROI per dollar. Prioritize size into this zone when signal quality is highest.

---

5. Both-sides allocation

The paired-spread mechanism is a secondary strategy, not the primary profit driver. Use it when:

1. Your forecast says one outcome is 90%+ likely
2. The near-certain outcome is priced at $0.97-$0.999 (confirming high liquidity)
3. Competing outcomes are priced at $0.001-$0.01

Allocation rules:

Near-certain winner:  80-90% of paired capital
Near-certain loser(s): 10-20% of paired capital (spread across all competing outcomes)

Target paired cost: below $0.995
Expected spread income: $0.005-$0.010 per share pair

Do not use the paired mechanism when you are uncertain about the outcome. The reference trader's 18/18 dominant-side win rate confirms he only pairs when he is already near-certain -- the pairing is a mechanism to extract the spread from a market he already knows the answer to, not a hedge against genuine uncertainty.

The paired mechanism's primary economic contribution is as a drawdown buffer: even in weeks when the directional longshot bets miss, the spread-capture P/L from near-certain outcomes provides steady small positive returns that prevent deep losses.

---

6. Bankroll math

Reference book 30-day performance:
  Buy notional deployed:       $30,409
  Trading P/L (resolved buys): +$5,420  (+17.4% ROI)
  Account total (Polymarket):  +$9,127  (+30% on approx working capital)

  P/L sources:
    Spread capture (18 paired markets): +$80 gross (offset by -$170 hedge tax)
    Near-certain winner settlements:    +$550  (2.5% ROI on $21,707)
    Longshot directional wins:          +$4,790 (from sub-$0.20 entries)

Monthly rate extrapolated:
  Working capital needed (peak):  ~$15,000-$25,000 (held in USDC on Polygon)
  Capital actually at risk daily:  ~$1,000-$3,000
  Expected monthly P/L:           +$4,000-$7,000 on $25,000 bankroll
  Expected monthly ROI on bankroll: +16-28%

Note: These returns are event-dependent and lumpy. In a month without a major
meteorological mispricing, expect +3-7% from spread capture and small longshot wins.
The outsized returns come from 1-3 major forecast calls per month.

---

7. Hour scheduling

Forecast publication timing: The key data inputs for this strategy are:

• 00:00 UTC NWP model run (GFS, ECMWF) -- available ~3-4 hours after run start
• 12:00 UTC NWP model run -- available ~3-4 hours after run start
• Meteorological agency public forecasts (China Meteorological Administration, Korea Meteorological Administration) -- published afternoon local time

Position entry should follow the data: wait for the most recent model run output, compute your probability estimate for each candidate temperature, identify markets where your estimate diverges 5x+ from the market price, then execute within 1-2 hours of receiving the data.

Do not trade while the forecast is stale. If you have not received updated model output in the last 8 hours, do not enter new positions. The market's crowd may have more recent information than you do.

---

8. Operational requirements

---

9. Risk profile

Maximum downside scenario: If 10 consecutive days of longshot bets all miss (historically plausible), the loss is bounded by the daily capital deployed in longshot positions -- typically $200-$500 per day, or $2,000-$5,000 maximum. The spread-capture positions (buying near-certain outcomes) continue generating +2-3% returns throughout, partially offsetting the longshot losses. There is no scenario where this strategy causes catastrophic loss, because each individual trade's maximum loss equals the premium paid (usually $0.01-$200 per position).

---

10. Diagnostic checklist -- is the strategy still working?

Run weekly:

---

11. What this playbook deliberately does NOT include

• No standard price filter ($0.30-$0.70). This destroys 93% of the trading profit. The alpha is at the extremes, not in the coin-flip zone.
• No dominance-only filter. The 100% dominant-side win rate is evidence of forecast quality, not a filter criterion. Using it as a filter keeps only the low-ROI near-certain positions and removes the high-ROI longshots.
• No crypto, sports, or political markets. The meteorological edge does not transfer to other categories. Each category requires its own edge source.
• No 24/7 continuous operation. This is a 2-3 hour per day strategy during the forecast data update window. Adding more trading hours without more data adds noise, not alpha.
• No uniform sizing. The power-law size structure (penny probes plus large conviction fills) is deliberate and essential. Switching to uniform sizing would cap your upside on the high-conviction events that generate most of the P/L.
• No diversification outside weather. The strategy's edge is meteorological knowledge. Diversifying into unrelated markets dilutes capital without adding edge.

---

TL;DR -- runnable summary

DAILY OPERATION (2-3 hours, 06:00-10:00 UTC):

1. Pull latest NWP model output for target cities:
   Shenzhen, Seoul, Hong Kong, Shanghai

2. For each city+date with active Polymarket markets:
   a. Compute forecast probability for each candidate temperature (24-48hr horizon)
   b. Compare to market prices

3. Entry criteria:
   LONGSHOT: forecast_prob >= 5 * market_price AND market_price <= $0.25
     - Entry size: $0.01-$0.50 (probe), $50-$500 (conviction)
     - Conviction threshold: forecast_prob >= 15% AND market_price <= $0.05
   
   SPREAD CAPTURE: forecast_prob >= 90% for one outcome
     - Buy winner at $0.97-$0.999 (80-90% of paired capital)
     - Buy all losers at $0.001-$0.01 (10-20% of paired capital)
     - Target paired cost below $0.995

4. Hold to resolution.
   Exit early only if new model run materially changes your probability estimate.

5. Weekly: track calibration (forecast prob vs actual hit rate).
   If calibration drift exceeds 10%, review data source.

Expected: +15-25% monthly ROI on deployed capital, lumpy.
Outsized months (like August 2026 reference) require 1-2 correct longshot calls.
Steady months generate +5-8% from spread capture and small wins.

The strategy is low-frequency, high-analysis. The execution is simple. The hard part is the meteorological model -- and that is the part that cannot be copied from the trade CSV alone.