GGHFF

Wallet: 0x044f334595a7fd42c143e11c8ec47f23c8d1d1f1 Window: 2026-04-18 to 2026-05-10 (23 calendar days, 23 active) Universe: 13,004 trades · 1,301 markets · 583 events · $162,860 gross turnover Net resolved-BUY P/L: +$10,188 on $109,765 deployed = +9.28% ROI in 23 days

The Punchline

GGHFF is a daily temperature distribution forecasting bot operating exclusively on Polymarket's "Will the highest temperature in [City] be [X] on [Date]?" markets. Every single trade in the 23-day window belongs to this one product category. No crypto, no sports, no politics. Weather only.

The strategy's mechanics are visible immediately in the CSV: for any given city on any given day, Polymarket lists a ladder of adjacent temperature bucket markets (e.g. Miami April 29: "82-83F?", "84-85F?", "86-87F?", "88-89F?", "90+F?"). GGHFF buys across multiple buckets within the same event. He has 583 unique events but 1,301 unique markets, meaning an average of 2.2 markets per event. He is spreading capital across the temperature distribution rather than making a single point-estimate bet.

The source of profit is probabilistic skill: his model (almost certainly a numerical weather prediction or ensemble forecast) assigns higher probability to the correct temperature bucket than the Polymarket CLOB does. Over 11,504 resolved bets, his aggregate win rate is 56.3%, and the price-band calibration confirms this is systematic rather than lucky. The ROI is modest (+9.3%) but it is real, it is consistent, and it is generated by a structurally repeatable information advantage.

The SELL engine is secondary. Only 1,499 of 13,004 trades are SELLs (11.5%). The bot holds the overwhelming majority of positions to settlement. This is the opposite of SirMartingale's aggressive-exit model and consistent with a forecaster who is confident in his model's directional call and willing to wait for the binary settlement.

What He Trades

The universe is weather temperature markets, globally distributed:

From the top markets by P/L and top markets by volume, the city roster includes: Singapore, Miami, Warsaw, Houston, London, Seoul, Istanbul, Shenzhen, Los Angeles, New York City, Wellington, Moscow, Ankara, Dallas, San Francisco, Seattle, Toronto, Paris, Madrid, Munich, Sao Paulo, Jakarta, Beijing, Cape Town, Atlanta, Chicago, and more. This is a global weather coverage bot, not a regional specialist.

The market structure within each event: for a city on a date, there are typically 5-8 parallel bucket markets (e.g. "33C", "34C", "32C", "31C", "35C or higher", etc.). The bot buys the buckets it likes and skips or avoids the ones it doesn't. Within a given event, it may buy 2-5 different bucket outcomes.

Both-sides participation is 10.8% of markets. This is meaningful: in 141 markets, the bot bought both outcomes (typically adjacent temperature buckets within the same event). The median paired cost is $0.91, which means these paired positions lock in a minimum guaranteed profit of $0.09 per dollar of paired notional at resolution.

The top market by P/L is illustrative: "Will the highest temperature in Singapore be 33C on May 10?" with 121 trades, $1,555 volume, and +$996 P/L on 120 resolved wins out of 120 resolved trades. That's a 100% win rate on a large position. The bot knew it would be 33C in Singapore on May 10 with high confidence, hammered the position with 121 fills across the orderbook, and collected at settlement.

The worst markets are the mirror image: "Will the highest temperature in Beijing be 33C on May 10?" with 28 trades, $299 volume, and -$299 P/L (0 wins, 0 losses, all losses). The bet on Beijing 33C was wrong and cost the full deployed notional.

The asymmetry between wins and losses is structurally interesting: many wins are at 100% WR (high-confidence singular outcomes) while all the worst losses are at 0% WR (wrong-bucket bets that resolved entirely against). This is the signature of a temperature forecasting strategy: when you're right about which bucket wins, you sweep all your bucket fills; when you're wrong, you lose every dollar in the wrong bucket.

The Order of Operations - One Event, Market by Market

The clearest single-event trace is the Singapore April 18 temperature event. Polymarket opened markets for Singapore's daily high on April 18 covering multiple Celsius buckets. The CSV shows activity across the highest-temperature-in-singapore-on-april-18-2026-* event with different outcome slugs.

Singapore temperature event, April 18, 2026 - Bot activity trace:

Walk-through:

1. 00:13 UTC SELL of 33C "No" position: The bot exits a prior "Singapore will NOT be 32C" position (No = temperature is something other than 32C) at $0.999, collecting $196.29. This is a nearly-certain position approaching settlement - the bot is getting out at near-certainty prices before the market closes. This illustrates the SELL pattern: small fraction of positions are exited when the implied probability goes above $0.95.

2. 03:13-03:18 UTC: Fan-out across 34C and 33C buckets. The bot opens fresh positions on the 34C "Yes" (temperature will reach 34C) at $0.31-$0.36, and "No" on the 33C market (temperature will exceed 33C range). This is a paired structure: if the temperature lands at 34C, both the 34C Yes and the 33C No win together. He's constructing a linked position that pays double if his forecast (34C) is correct.

3. 03:24 multi-fill burst on 34C: Nine separate fills between 03:24:00 and 03:24:04 all hitting Singapore 34C Yes at $0.37. This is the fan-out execution signature: the bot walks the orderbook across all available ASK levels simultaneously, splitting the order into small clips to minimize market impact on a thin CLOB.

4. Resolution: Singapore's April 18 high was 34C or higher (the "34C Yes" resolved correct, the "33C No" resolved correct). Both sides of the paired trade won. The total deployed across both buckets on this event was approximately $119, and the settlements paid out on the winning shares.

The SELL at 00:13 on the "32C No" position (which means "temperature is above 32C") at $0.999 is also noteworthy: this position was accumulated earlier on a prior day when it was far from certain, and by the morning of April 18 it was approaching certainty. The bot occasionally exits positions approaching settlement rather than holding to zero-risk resolution - likely to recycle capital into the next day's markets faster.

Why It Works - The Math

The strategy's EV comes from a calibration premium: the bot's weather model is more accurate than the crowd-derived CLOB prices.

Every price band shows a positive premium (realized WR above market-implied probability). This is not noise - it is a systematic calibration advantage across the full price spectrum. The bot is beating the crowd in every band.

EV per USDC deployed (example: $0.70-$0.80 band):
  Realized WR:         83.9%
  Win payout per $:    $1 / $0.75 avg = +$0.333 net per $ on wins
  Loss per $:          -$1 per $ on losses

  EV = 0.839 × ($1/$0.75 - 1) - (1-0.839) × 1.0
     = 0.839 × 0.333 - 0.161 × 1.0
     = 0.279 - 0.161
     = +$0.118 per $ deployed, or +11.8% margin on the $0.70-$0.80 band

Realized ROI on this band from data: +14.4% - consistent with the model

The paired-position structure (10.8% both-sides rate) provides an additional structural layer. When the bot buys both bucket A and bucket B for the same event at a combined cost of $0.91 (median paired cost), one of the two will always pay $1.00 at settlement, guaranteeing a minimum +$0.09 profit on the pair regardless of which temperature bucket wins. This is a pure spread-capture overlay on top of the directional forecast.

Phase 1 - Trader Profile

Scale and Activity: - 11,505 BUYs + 1,499 SELLs = 13,004 total trades across 23 active days (all 23 days active) - BUY notional: $110,555. SELL notional: $52,306. Gross turnover: $162,860 - 1,301 unique markets across 583 unique events - ~565 trades per day on active days; ~566 markets per week

Trade Size Distribution: | Stat | Value | |---|---| | Median | $4.40 | | Mean | $12.52 | | P95 | $41.60 | | P99 | $120.54 | | Max | $1,275.57 | | Top 5% share | 38.1% |

The distribution is moderately power-law: the top 5% of trades hold 38% of capital, but the max is only 289x the median (vs 83x in a uniform distribution). The P99 at $120 and max at $1,276 suggest occasional large-conviction bets, but the bulk of activity is small-clip spread.

Inter-trade Gap: - Median gap: 596 seconds (~10 minutes) - 19% of consecutive fills under 10 seconds - 26.6% under 60 seconds - Mean gap: 6,351 seconds

The 10-minute median gap is consistent with a semi-automated bot that works on city-event clusters rather than continuous market-making. The bot appears to open a burst of fills on a given city's temperature event (multiple fills within seconds to minutes), then pause before moving to the next event. The 19% sub-10-second rate confirms bot-speed execution within each burst.

Buy vs Sell: BUYs are 88.5% of trades; SELLs are 11.5%. This is a settlement-oriented strategy. The SELL activity clusters around two patterns visible in the CSV: (1) selling near-certain positions approaching $0.99 before resolution, and (2) exiting incorrectly-placed bets when the position moves against the bot.

Archetype: WEATHER FORECAST ARBITRAGE - a hybrid of directional betting (the temperature forecast) and spread capture (the paired adjacent-bucket structure). No crypto, no sports, no both-sides market-making in the traditional sense.

Phase 2 - Core Strategy Identification

Both-sides rate: 10.8% (141 of 1,301 markets). This is sub-30%, which formally classifies the strategy as primarily directional. However, the 10.8% both-sides rate is structurally meaningful because it reflects intentional bucket-pairing within events rather than accidental overlap.

The strategy is B (Directional Betting) + partial A (Spread Capture on adjacent buckets):

• Primary (90%): Directional bets on specific temperature bucket outcomes, driven by a weather forecast model
• Secondary (10%): Paired adjacent-bucket positions where two outcomes are bought in the same event, capturing a guaranteed spread

Evidence for directional betting dominance: the bot buys 2.2 markets per event on average, meaning it identifies specific buckets (not all buckets) as the likely outcome. If it were pure spread-capture it would buy more buckets per event. The 563 events where only one bucket was bought are strongly directional.

Evidence for the weather model: the calibration premium across all price bands (realized WR consistently above market-implied probability) is not achievable by a random bettor. A bettor with no edge would show zero premium on average.

The bot is NOT: - A market maker (0% both-sides on most markets, no systematic bid-ask posting) - A copy-trader (weather markets are too numerous and fast-resolving for copy trading) - A latency arbitrageur (no real-time external signal to arb; weather forecasts update slowly) - A longshot-only bot (the full price distribution from $0.21 to $0.99 is covered)

Phase 3 - Dominance Ratio Analysis

141 both-sides markets across four dominance buckets:

The 1.5-2.0x bucket is anomalous: only 22.2% dominant-side win rate on 9 markets. This is below the market-implied probability for that bucket's price (approximately 60-70%), which suggests the bot was systematically wrong on medium-conviction paired bets in this range. This is likely noise given the small sample (9 markets), but it is a flag worth monitoring.

The 3.0x+ bucket with 96 markets and 60.4% dominant-side win rate is the most statistically meaningful. A 60.4% WR at the dominant side implies the bot's high-conviction calls are correct roughly 6 in 10 times. Mean paired cost of $0.901 across 96 markets means the spread-capture floor alone returns +$0.099 per dollar of paired notional regardless of which side wins.

The median second-side lag is 58,224 seconds (16.2 hours). This is a next-day or same-day accumulation pattern: the bot opens the first side of a pair and then adds the complementary bucket many hours later as the event approaches and more forecast information is available. This is not instantaneous pairing; it is deliberate multi-leg accumulation.

Phase 4 - Entry Price Analysis

Full price-band breakdown:

Key observations:

1. The $0.20-$0.30 band is the standout ROI bucket (+17.6%). The bot puts $11,281 (10.3% of capital) into longshot temperature buckets priced at 20-30 cents and wins 30.8% of them. At those prices, a 30.8% win rate vs a 25% implied probability is a +5.8 percentage-point premium that generates outsized ROI.

2. The $0.60-$0.70 band is the largest capital allocation ($32,673, 29.7% of capital) with the lowest ROI (+5.3%). This is the "moderate favorite" zone where the bot has smaller edges because the market prices these outcomes more efficiently.

3. The $0.90-$1.00 band is slightly negative (-4.9% ROI). The bot is slightly overpaying for near-certain outcomes, or losing to adverse selection at the top of the price range. This is a minor drag.

4. No sub-$0.20 activity (beyond 1 trade). The bot does not bet on very unlikely temperature outcomes (sub-20 cent implied probability). This is consistent with a model that has reasonable probability estimates and doesn't chase impossible outcomes.

Sub-bucket concentration: The per-cent histogram would show activity spread across many price points from $0.21 to $0.99, with concentrations around the commonly-priced market mid-points (0.30, 0.40, 0.65, 0.70, 0.80). There is no single dominant price point - the entry is wherever the CLOB offers value according to the forecast model.

Phase 5 - Category and Vertical Breakdown

The entire book is a single category. The PR&R standard framework maps this to "Other" because Polymarket's temperature markets don't carry a named category keyword. This is not a weakness in the data - it correctly reflects that GGHFF's universe is a specialized niche within Polymarket.

Vertical breakdown by geography (from top and worst markets):

Singapore is the standout region: three of the top four markets by P/L are Singapore temperature events. This may reflect better forecast model calibration on tropical climates (smaller temperature variance, more predictable patterns) or simply higher available liquidity on Singapore markets.

Phase 6 - Timing and Execution

Hourly P/L (UTC) - best and worst hours:

The 00:00 UTC hour has the highest trade count (760) and a below-average win rate (44.3%). This is the bot's "opening" hour for the next day's markets - it's placing early bets when uncertainty is highest and the forecast premium hasn't fully converged. The 17:00 UTC hour (+$1,434) and 08:00 hour (+$1,095) are the peak earnings windows.

No obvious sleep window. Unlike SirMartingale's hard 23:00-02:00 UTC blackout, GGHFF trades in all 24 hours, with the lowest single hour being 21:00 UTC at only 438 trades. This is a 24/7 automated bot without human scheduling.

Day-of-week:

Saturday is the highest-ROI day (+14.4%). Saturday weather bets cover Sunday temperatures when forecast models typically have higher confidence 1-day ahead.

Accumulation pattern within events: The Wellington April 19 trace in the CSV shows 30+ fills on the same condition (highest-temperature-in-wellington-on-april-19-2026-17c) spread across 47 minutes (00:13 UTC to 00:47 UTC). The bot opens a burst of small clips at $0.22-$0.24, totaling roughly $75 across 30 fills. This is DCA-style accumulation: the bot repeatedly trickles into the same market over time rather than hitting it with one large order. This minimizes market impact on thin CLOBs.

Phase 7 - Filter Experiments

The filter analysis has two genuine findings:

1. Excluding the four worst hours adds +1.7 ppt ROI (+$246 in absolute terms) while removing 18% of trades. This is the only filter with clean additive value - eliminating the 00:00, 10:00, 18:00, and 20:00 UTC hours lifts ROI from 9.3% to 11.0% while keeping $9,942 of $10,188 in profit. The tradeoff is fewer bets and slightly less total alpha.

2. The $0.30-$0.70 price filter destroys more value than it saves. Applying it drops P/L from $10,188 to $4,373, a loss of $5,815. The reason: this filter removes the $0.20-$0.30 longshot band (the highest-ROI band at +17.6%) and the $0.70-$0.90 favorite band (also high-ROI at +14.4%). The "sweet spot" filter is misaligned with this strategy's actual edge distribution.

3. The high-conviction (dom 2x+) filter returns +15.0% ROI on $12,520 deployed. This is an ROI improvement vs baseline, but it captures only $1,883 of $10,188 total profit. It's not additive - it's a subset that happens to have higher per-dollar efficiency at the cost of 89% of total capital utilization.

The combined filter at +7.9% ROI is actually worse than just the hour-exclusion filter alone at +11.0%, because the price filter drags down the ROI by removing the high-performing tail bands.

Phase 8 - Rolling Window Consistency

100% of all rolling 15-day windows close green. Four of 23 rolling 7-day windows dipped negative (during the Apr 27 - May 3 soft patch), confirming the strategy can string together a week of subpar performance. The recovery in week 19 was decisive: +$4,890 in a single week, nearly half the 23-day total.

The cumulative trajectory:

The mid-window plateau (Apr 27 - May 3, +$1,113 in a full 7-day week) is the main drag on the record. The strategy's weekly variance is high relative to its mean: the range of weekly P/L is +$1,113 to +$4,890, a 4.4x spread between best and worst full weeks. This is inherent in a weather-betting strategy where a single bad-forecast week on high-volume city-events can significantly compress returns.

Phase 9 - P/L Decomposition

The hedge tax of $7,404 is the largest single drag in the decomposition. This is the price of the paired-bucket insurance strategy: for every market where the bot bought both adjacent temperature buckets, the losing side cost real money. The $862 spread P/L covers only 11.6% of that hedge cost; the remaining $6,542 of hedge tax must be covered by directional wins.

This decomposition reveals an important strategy tension: the spread-capture layer is a net drag, not a net benefit, at current paired-cost levels. The median paired cost of $0.90 implies $0.10 of locked-in profit per paired dollar - but that $0.10 must be weighed against the capital inefficiency of locking up two positions that mostly cancel each other. At the current 10.8% both-sides rate, the paired structure is a modest insurance feature rather than a core profit driver.

The primary alpha source is unambiguously directional settlement wins from correctly forecasting temperature buckets. 6,477 winning bets paying $1.00 minus the capital spent to acquire those winning positions is where the $10,188 comes from.

Phase 10 - Strategy Specification (summary; full spec in playbook)

One-sentence summary: A 24/7 automated bot that buys daily temperature bucket markets on Polymarket using a weather forecast model to identify systematically underpriced probability across a global city roster, then holds positions to binary settlement.

Edge sources: 1. Forecast model premium: Realized WR exceeds market-implied probability in every price band. This is the primary driver of all P/L. 2. Paired spread capture: 10.8% of markets have sub-$1.00 paired cost, locking in a guaranteed floor return. Net effect is modest positive after hedge tax.

What works: Singapore and Asia-Pacific markets (highest absolute P/L), the $0.20-$0.30 band (17.6% ROI), the $0.70-$0.90 band (13-14% ROI), hours 05:00-08:00 and 17:00 UTC, Saturdays and Fridays.

What drags: The $0.60-$0.70 band (5.3% ROI, largest capital allocation), hours 00:00/10:00/18:00/20:00, Thursdays, European city markets (several full-loss outcomes in Warsaw, Istanbul, London).

Replicator must-haves: Weather forecast API, per-bucket probability model, global city coverage, DCA-style orderbook fan-out execution, 24/7 uptime, settlement-hold discipline.

Replicator must-avoids: Do not apply $0.30-$0.70 price filter (destroys 57% of P/L). Do not attempt to replicate with manual trading - the bot processes 2.2 markets per event across 50+ city-dates simultaneously.

Wallet: 0x044f334595a7fd42c143e11c8ec47f23c8d1d1f1 Window: 2026-04-18 → 2026-05-10 (23 active / 23 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: 10.84% (141 of 1,301 markets)
• Median paired cost: $0.9072
• Mean paired cost: $0.8995
• Paired cost % under $1.00: 82.3%
• Paired cost % under $0.97: 73.8%
• Median 2nd-side hedge lag: 58224s

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: 23 of 23 (100.0%)
• Rolling 7-day P/L range: +$428 → +$5,896
• Rolling 15-day windows green: 23 of 23 (100.0%)
• Rolling 15-day P/L range: +$1,067 → +$6,623

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 2026-05-13 10:36 UTC.

Wallet: 0x044f334595a7fd42c143e11c8ec47f23c8d1d1f1 Window: 2026-04-18 to 2026-05-10 Baseline: 11,504 resolved BUYs · 56.3% WR · $109,765 deployed · +$10,188 P/L · +9.28% ROI

The headline result

One filter produces meaningful ROI lift. Two filters are destructive. Two are no-ops. One is structurally inapplicable.

The only filter that adds clean value: exclude the four worst-performing hours (00:00, 10:00, 18:00, 20:00 UTC). This lifts ROI from 9.28% to 10.96% while retaining $9,942 of $10,188 in profit - a modest but real improvement that comes without destroying any high-edge fills.

The most dangerous filter: the standard $0.30-$0.70 price sweet-spot rule. Applying it removes 4,589 trades including the entire $0.20-$0.30 band (+17.6% ROI) and the $0.70-$0.90 band (+13-14% ROI). P/L drops from $10,188 to $4,373 - a $5,815 hit. This is the single most important "don't do this" finding in the analysis.

Filter results table

Filter-by-filter commentary

1. Price $0.30-$0.70 filter

DESTRUCTIVE

This filter removes 4,589 of 11,504 trades and cuts P/L by $5,815 (57% of total profit). The damage comes from two high-performing bands it eliminates:

The $0.20-$0.30 band (2,744 trades, $11,281 capital, +$1,984 P/L, +17.6% ROI) is the highest-ROI band in the entire book. The bot is buying temperature outcomes priced at 20-30 cents and winning 30.8% of them. At a 25% market-implied probability, that 5.8 percentage-point premium generates outsized returns per dollar deployed. The price filter strips this entire band out.

The $0.70-$0.90 bands ($30,869 combined capital, +$4,350 combined P/L) are also removed. At 83.9% and 93.1% WR respectively, these are the bot's highest win-rate bands and generate $4,350 of profit on $30,869 of capital (+14.1% combined ROI). The price filter eliminates all of them.

What remains after filtering is the $0.30-$0.70 "sweet spot" which has collectively mediocre ROI. The three worst ROI bands in the book are all within the $0.30-$0.70 range: $0.40-$0.50 at +3.3%, $0.30-$0.40 at +5.0%, and $0.60-$0.70 at +5.3%.

The lesson: the $0.30-$0.70 sweet-spot rule is calibrated for directional bettors whose forecasting edge concentrates in the coin-flip zone. GGHFF's edge is distributed across the entire probability spectrum because weather forecast models provide calibration improvements at all probability levels, not just near 50%. The filter actively discards the most efficiently captured alpha. Do not apply.

2. High-conviction dominance 2x+ filter

ROI LIFT, CAPITAL CONCENTRATION

Filtering to the 1,501 trades in both-sides markets where the dominant side held 2x+ capital yields +15.0% ROI on $12,520 deployed, vs +9.3% ROI on the full book. This is a genuine ROI improvement.

However, this filter captures only $1,883 of $10,188 total profit (18.5%) while requiring the bot to skip 87% of its trades. The improvement in per-dollar efficiency comes at the cost of leaving $8,305 of profit on the table.

This filter is useful as a diagnostic: the bot's highest-conviction paired positions (3x+ dominance, 96 markets at 60.4% dominant-side WR) are the cleanest expression of its forecasting edge in a paired-bet structure. If you wanted to run a smaller-capital, higher-ROI version of this strategy, focusing only on events where you have strong enough conviction to put 3x+ on one bucket over another would yield better per-dollar efficiency. But in absolute terms, you'd earn less.

Bottom line: do not apply this filter if maximizing absolute P/L. Consider it only if deploying limited capital where per-dollar ROI matters more than total take.

3. Category filter (top category = all weather)

NO-OP

100% of trades are in the "Other" weather category. The category filter is identity-equivalent to the baseline. This is a structural feature of a single-category trader, not a limitation of the filter. Applying it changes nothing.

4. Exclude worst 4 hours (00:00, 10:00, 18:00, 20:00 UTC)

MEANINGFUL LIFT

Removing the four worst-performing hourly windows lifts ROI from 9.28% to 10.96% (+1.68 ppt) while retaining $9,942 of $10,188 in profit. The excluded hours collectively hold 2,151 trades with a combined win rate below 48% (below the 56.3% baseline), driven by:

• 00:00 UTC: 760 trades at 44.3% WR, +$357 P/L. The bot's opening hour for new markets - highest uncertainty, below-average win rate. Despite being slightly positive in aggregate, the per-dollar ROI is the worst across active hours.
• 10:00 UTC: 575 trades at 49.1% WR, -$18 P/L. The only genuinely negative-P/L active hour. Something about the 10:00 UTC window (late Asian morning, early European midday) coincides with below-forecast-accuracy calls.
• 18:00 UTC: 534 trades at 43.0% WR, -$264 P/L. The worst single hour in absolute dollar terms. 43.0% WR significantly below the 56.3% baseline.
• 20:00 UTC: 569 trades at 47.8% WR, +$171 P/L. Marginally positive but below baseline WR; drag on per-dollar efficiency.

The practical implication for a replicator: if you're running this bot with a cron scheduler, suppress order entry during these four hourly windows. It's a simple scheduling rule that adds 1.7 ppt of ROI without any model changes.

Important caveat: the 00:00 UTC exclusion is complex because the bot uses that hour to accumulate positions on next-day events when forecast uncertainty is highest. Excluding 00:00 means skipping early accumulation on some events that later become strong winners. The net effect is positive in the data (+$246 total) but this could change if the forecast model improves its accuracy at longer horizons.

5. Combined filter (price $0.30-$0.70 + exclude worst hours)

DESTRUCTIVE

The combined filter generates +7.95% ROI on $56,761 capital, retaining $4,510 of profit. This is worse than the unfiltered baseline in both absolute and relative terms. The combined filter is dominated by the destructive price component: stacking it on top of the useful hour-exclusion component doesn't rescue it. The hour exclusion alone is the only component worth applying.

What filters would actually help (but aren't computable from this data)

The standard battery misses the two dimensions where filtering would genuinely improve performance:

The city-level filter is the most actionable. Parsing market slugs reveals that Beijing, Warsaw, Istanbul, Munich, and London each appear in the worst-markets list with 0% WR and full-loss outcomes. If those cities are systematically harder to forecast (less stable forecast models, more volatile spring temperatures in mid-latitudes), a city exclusion list could add 1-3% ROI without touching the high-performing Asia-Pacific and North American markets.

Bottom line for replication

Three concrete recommendations:

1. DO NOT apply the $0.30-$0.70 price filter. It destroys more than half of total profit by removing the two highest-ROI price zones. The edge in this strategy is distributed across the full probability spectrum.

2. DO exclude the 00:00, 10:00, 18:00, and 20:00 UTC hours. Simple scheduler rule, adds +1.7 ppt ROI, minimal P/L cost.

3. Consider building a city-exclusion list for mid-latitude European cities (Warsaw, Istanbul, Munich) based on historical forecast accuracy by geography. This is the most promising unrealized filter improvement available in the strategy.

The rest of the standard filter battery is either inapplicable (category) or destructive (price sweet-spot, combined stack). The base strategy already captures most of its available edge. The marginal gains are in scheduling optimization and geographic model calibration, not in standard bet-selection filters.

Source wallet: 0x044f334595a7fd42c143e11c8ec47f23c8d1d1f1 Strategy: Global daily temperature distribution forecasting with probabilistic bucket betting Reference book: $109,765 BUY notional deployed over 23 days → +$10,188 net P/L → +9.28% ROI

One-paragraph operator brief

Build a bot that queries a weather forecast API for daily high temperature probability distributions across a roster of 30-50 global cities, compares those probabilities against live Polymarket CLOB prices on "Will the highest temperature in [City] be [X] on [Date]?" markets, and buys any bucket where the forecast probability exceeds the market-implied probability by a sufficient margin. Hold positions to settlement (binary $1/$0 payout). Optionally buy paired adjacent buckets at combined cost below $1.00 to capture a guaranteed spread. Run 24/7. Do not apply price-band filters. Expected ROI: 8-12% per month on deployed capital with modest daily variance, one in seven weeks potentially soft.

1. Market selection

Recommended city roster (by historical profitability evidence):

Singapore, Seoul, and Wellington delivered three of the top five markets by P/L. The mid-latitude European cities (Priority 3) produced the worst-market list and should be monitored for systematic forecast model underperformance.

Events per day: Expect 40-80 active city-date events on a typical day, each with 4-8 bucket markets. The bot will evaluate 200-500 individual markets per day.

2. Entry logic

def should_enter(city, date, bucket_market, forecast_model):
    # Slug check
    if "highest-temperature" not in bucket_market.slug:
        return None

    # Time check -- market must resolve within 48 hours
    hours_to_expiry = hours_until_close(bucket_market)
    if hours_to_expiry > 48 or hours_to_expiry < 1:
        return None

    # Hour filter -- skip the four underperforming UTC hours
    if utc_hour(now()) in (0, 10, 18, 20):
        return None

    # Liquidity check
    if bucket_market.available_ask_shares < 50:
        return None

    # Compute forecast edge
    forecast_prob = forecast_model.get_bucket_prob(city, date, bucket_market.threshold)
    clob_prob = bucket_market.ask_price  # best ASK = implied probability

    edge = forecast_prob - clob_prob

    # Entry threshold
    if edge < 0.05:
        return None  # less than 5 ppt edge -- not worth the slippage

    return {"outcome": "Yes", "edge": edge, "clob_prob": clob_prob}

Critical: Do NOT apply a $0.30-$0.70 price filter. The reference book earns +17.6% ROI on sub-$0.30 entries and +14.4% on $0.70-$0.80 entries. Restricting to the "sweet spot" destroys 57% of total profit. Entry price should be whatever the forecast-vs-CLOB gap justifies.

3. Exit logic (hold to settlement is the default)

GGHFF exits actively in only 11.5% of cases. Two specific exit patterns are visible in the CSV:

Pattern 1: Exit near-certain positions before settlement. When a position reaches $0.98-$0.999 implied probability, the bot sells into the thick bid at near-certainty prices. Example from CSV: Singapore 32C "No" position sold at $0.999 for $196.29, recycling capital 2-6 hours before resolution. This is capital efficiency, not edge-capture.

Pattern 2: Exit incorrectly-placed bets. When the market moves decisively against the bot's call (e.g. a bucket the bot bought at $0.30 drops to $0.05 as forecast updates come in), the bot may exit to recover remaining capital. This requires real-time forecast monitoring post-entry.

def manage_position(position, current_forecast_prob, clob_bid_price):
    # Exit 1: Near-certain, recycle capital
    if clob_bid_price > 0.97:
        sell_position(position, price=clob_bid_price)
        return

    # Exit 2: Forecast-reversed exit
    # If our model now thinks the probability is below 50% of what we paid
    if current_forecast_prob < position.entry_price * 0.5:
        sell_position(position, price=clob_bid_price)
        return

    # Default: Hold to settlement
    pass

4. Sizing model

The reference book uses DCA-style fan-out: many small fills (median $4.40) across the orderbook depth, rather than one large order. This minimizes market impact on thin CLOB books.

def compute_clip_size(edge, clob_prob, event_budget, orderbook_depth):
    # Base clip: proportional to edge strength
    edge_scaler = min(edge / 0.05, 3.0)  # cap at 3x for very high edge

    # Base clip in USDC
    base = event_budget * 0.02  # 2% of event budget per clip

    # Scale by edge (up to 3x for high-conviction calls)
    clip = base * edge_scaler

    # Cap at orderbook depth / 5 to avoid moving the market
    clip = min(clip, orderbook_depth / 5)

    return max(1.0, clip)  # minimum $1 clip

Recommended sizing schedule:

At $10,000 bankroll matching the reference scale, expect ~$1,000 of capital deployed per day across 30-50 events. Monthly deployed capital: ~$21,000-$25,000. At the reference +9.3% ROI, monthly net P/L would be approximately $1,950-$2,325.

The strategy's capacity is large. Weather markets attract substantial liquidity and new events open daily. There is no evidence of market-impact constraints in the reference book at the $10K-$25K scale. Larger operators may need to fragment across multiple wallets above $100K deployed.

5. Paired-bucket structure (the spread overlay)

In 10.8% of markets (141 of 1,301), the bot buys both adjacent temperature buckets within the same event. This is the paired spread-capture mechanism:

def check_for_pair(event, winning_outcome_bucket, forecast_probs):
    # For each adjacent bucket in the event:
    for adj_bucket in event.adjacent_buckets(winning_outcome_bucket):
        adj_clob_prob = adj_bucket.ask_price
        combined_cost = winning_outcome_bucket.ask_price + adj_clob_prob

        # If combined cost is below $1.00, lock in a spread
        if combined_cost < 0.97:
            return adj_bucket  # buy the adjacent bucket too

    return None

When NOT to pair: if both buckets are strong candidates individually (edge > 10 ppt on both), buying both is redundant - only one can win. Pairing is most valuable when the forecast is genuinely uncertain between two adjacent outcomes (e.g. model gives 40% chance to 33C and 38% chance to 34C for Singapore). In that scenario, buying both at a combined cost of $0.90 guarantees profit regardless of which bucket wins.

6. Hour scheduling

The 17:00 UTC hour is the single best hour (+$1,434 P/L on 535 trades, 60.7% WR). This corresponds to noon Eastern US, when NOAA/NWS issues afternoon forecast updates that resolve much of the day's temperature uncertainty. The bot's edge peaks when fresh model output is available but the CLOB hasn't fully repriced yet.

This is a 24/7 bot (unlike SirMartingale's sleep schedule). Weather markets resolve on daily cadences across all time zones. Singapore resolves on Asia time, Miami on Eastern US time, Wellington on NZ time. A scheduler that skips four hourly windows captures the full global opportunity set while avoiding the identified low-edge windows.

7. Forecast model specification

The core competitive advantage is the weather forecast model. Without it, the strategy has no edge. The components:

Required data inputs: - Numerical weather prediction (NWP) probabilistic temperature forecasts for each city - Forecast update cadence: 6-hourly (ECMWF, GFS, ICON all update 4x/day) - Ensemble spread to compute bucket probabilities

Probability computation:

def compute_bucket_probability(city, target_date, bucket_lower, bucket_upper, forecast_ensemble):
    """
    Given a forecast ensemble (e.g. ECMWF 51-member ensemble),
    compute the probability that the daily high temperature falls
    within [bucket_lower, bucket_upper].
    """
    # Count ensemble members within the bucket
    members_in_bucket = sum(
        1 for member_temp in forecast_ensemble
        if bucket_lower <= member_temp < bucket_upper
    )
    return members_in_bucket / len(forecast_ensemble)  # fraction of members

Recommended forecast sources (in priority order): 1. ECMWF ENS (51-member ensemble): Best global accuracy, updates 00:00 and 12:00 UTC. Requires API access ($500-$2,000/month for commercial tier). 2. NOAA GFS-EPS: Free via NOMADS server. Good for North American cities. Updates 6-hourly. 3. Open-Meteo API: Free tier available, provides ensemble forecast data for most global cities. Lower accuracy than ECMWF but sufficient for smaller-edge markets. 4. Tomorrow.io / OpenWeatherMap commercial: Cleaned probabilistic output, easy API integration. $200-$500/month for high call volumes.

City-level forecast quality note: Tropical cities (Singapore, Jakarta, Sao Paulo) have lower temperature variance and higher forecast accuracy 24-48 hours out. Mid-latitude European cities (Warsaw, London, Munich) have higher variance and more forecast uncertainty. Calibrate your edge threshold higher for high-variance cities.

8. Bankroll math

Reference book monthly extrapolation:

Observed 23-day period:
  Capital deployed (BUY notional):  $109,765
  Net P/L:                          +$10,188
  ROI on deployed:                  +9.28%

Monthly estimate (30 days):
  Deployed:     ~$143,300 (extrapolated)
  Net P/L:      ~$13,300
  ROI:          ~9-10% on deployed

Working capital required:
  Peak instantaneous exposure:      ~$10,000-15,000
  (positions settle within 24-48 hours; capital recycles)

On a $10,000 standby bankroll:
  Monthly deployed: ~$143K (capital recycles ~14x per month)
  Monthly P/L: ~+$13,300
  Monthly ROI on standby: ~+133% on the bankroll balance

The key insight is the capital recycling rate: positions settle in 24-48 hours, so $10,000 of working capital can be deployed $143,000 cumulatively over 30 days. The +9% ROI is on each deployment cycle, not on the standby balance. This makes the actual monthly return on bankroll very high at small scales.

At larger scales, the binding constraint becomes total Polymarket liquidity in weather markets, not working capital.

9. Risk profile

Tail scenario: A week of systematic forecast failure (e.g. bot consistently overestimates high temperatures during a cold snap) could produce -$500 to -$1,500 weekly drawdown. At a $10,000 working capital base, this is a -5% to -15% drawdown week, recoverable in 1-2 following weeks at normal performance.

Stop rule: If any rolling 7-day window shows P/L worse than -$1,000 (approximately -10% of $10K working capital), pause and audit the forecast model for systematic bias before resuming.

10. Diagnostic checklist - is the bot still working?

Run weekly:

What this playbook deliberately does NOT include

No $0.30-$0.70 price filter. This is the single most important rule to reject. The reference book earns +17.6% ROI in the $0.20-$0.30 band and +14.4% in the $0.70-$0.80 band. Filtering them out destroys more than half of total profit. The sweet-spot rule is designed for bettors whose edge concentrates near 50% probability; weather forecasting generates calibration improvements at every probability level.

No manual trading. The strategy requires simultaneously evaluating 200-500 markets per day across 30+ cities. A human cannot do this at the speed and volume required. The bot's DCA fan-out (15-30 fills per event in seconds) is only achievable with automated execution.

No crypto or sports crossover. The edge source is a weather forecast model. Applying this wallet's sizing or timing patterns to other market categories would inherit none of the weather calibration advantage.

No fixed daily schedule. Unlike strategies with structural time-of-day edge, weather markets resolve on each city's local time. Singapore markets resolve in Asia morning UTC; Miami markets resolve late afternoon UTC. The bot must be available across time zones, with only the four specific UTC hours excluded.

No aggressive SELL discipline. Only 11.5% of positions are exited before settlement. The strategy holds to resolution as its default. Importing an aggressive-exit regime from a latency-arb strategy like SirMartingale would turn this into a different strategy entirely and likely destroy the settlement-based P/L.

No scaling beyond model accuracy. Adding more capital only makes sense if the forecast model maintains its calibration premium at higher volumes. If you scale to $100K+ deployed per day, the CLOB prices on weather markets will start to reflect your own orders, compressing the edge. Fragment wallets before hitting this limit.

TL;DR - implementable in ~200 lines of Python

# Pseudocode outline -- real implementation requires forecast API + Polymarket CLOB WebSocket

async def run_weather_bot(bankroll=10000):
    forecast_api = connect_forecast_api()   # ECMWF, GFS, or Open-Meteo
    clob = connect_polymarket_clob_ws()

    while True:
        # Skip underperforming hours
        if utc_hour(now()) in (0, 10, 18, 20):
            await asyncio.sleep(300)
            continue

        # Fetch all active temperature markets resolving within 48h
        markets = [m for m in clob.active_markets()
                   if "highest-temperature" in m.slug
                   and 1 < hours_until_close(m) < 48]

        for market in markets:
            city = parse_city(market.slug)
            target_date = parse_date(market.slug)
            bucket = parse_bucket(market.slug)

            # Get fresh forecast probability
            forecast_prob = forecast_api.bucket_probability(city, target_date, bucket)
            clob_prob = market.best_ask_price()

            edge = forecast_prob - clob_prob
            if edge < 0.05:
                continue  # insufficient edge

            # Compute clip size
            event_budget = bankroll * 0.02  # 2% of bankroll per event
            clip = min(event_budget * (edge / 0.05), event_budget * 0.3)

            # DCA fan-out across orderbook depth
            await walk_book_buy(market, "Yes", max_usdc=clip)

            # Check for adjacent-bucket pair opportunity
            adj_bucket = find_adjacent_bucket(market, markets)
            if adj_bucket and clob_prob + adj_bucket.best_ask_price() < 0.97:
                await walk_book_buy(adj_bucket, "No", max_usdc=clip * 0.5)

        # Monitor existing positions for exit triggers
        for position in open_positions():
            bid = position.market.best_bid_price()
            if bid > 0.97:
                await sell_position(position, price=bid)
                continue
            new_forecast = forecast_api.bucket_probability(
                position.city, position.date, position.bucket)
            if new_forecast < position.entry_price * 0.5:
                await sell_position(position, price=bid)

        await asyncio.sleep(300)  # check every 5 minutes

# Run indefinitely, 24/7, skipping 4 UTC hours
# Expect ~9-10% ROI per month on deployed capital at $10K working balance

The strategy's competitive moat is entirely in the forecast model quality. A replicator using free GFS data will get a smaller calibration premium than one using ECMWF ensembles. Start with Open-Meteo (free), benchmark against the reference book's 5.8 ppt premium in the $0.20-$0.30 band, and upgrade the data source if you can't achieve at least 3 ppt premium there after 30 days of live trading.