BillBenter2026

Wallet: 0x84ad9c5c547a82ec9a08547b94bd922446e5bfb7 Window: 2026-04-11 → 2026-05-07 (27 calendar days, 27 active) Universe: 892,903 trades · 10,243 markets · 4,521 events · $14.88M gross turnover Net P/L: +$96,765 on $14.87M deployed = +0.65% ROI in 27 days

The Punchline

This is not a pure market maker, and it is not a pure directional bettor. It is both, stacked. The bot buys both sides of the same event on 79.1% of its active markets, which superficially resembles market making. But the paired cost across the full book averages $1.033 - above the $1.00 break-even line - so the spread mechanic is losing money, not printing it. Every dollar of profit flows from the directional tilt: when the bot has conviction, it puts 3x or more on the dominant side, and that dominant side wins 73.0% of the time.

The dominance curve is the single most important finding. Win rates escalate monotonically with conviction: 54.4% at 1-1.5x, 58.0% at 1.5-2x, 61.1% at 2-3x, 73.0% at 3x+. The high-conviction bin alone contains 4,505 markets and generates the bulk of net positive P/L. This is a calibrated model expressing directional views through asymmetric allocation within paired positions. It is exactly what Bill Benter did in Hong Kong horse racing - a model-driven directional overlay on top of a structural participation mechanic.

The scale is extraordinary: 892,903 fills in 27 days, ~33,000 fills per day, active across all 24 UTC hours, spanning Soccer, Tennis, NBA, MLB, NHL, NFL, and esports. The low headline ROI (+0.65%) reflects the capital intensity of running 380+ markets per day. Absolute P/L of +$96,765 in 27 days on a book that turns over $14.9M is the real metric.

What He Trades

The universe is deliberately broad across sports categories, with Tennis and Soccer as the volume backbone:

Soccer is the largest category by trade count and is the biggest absolute P/L drag (-$58,634). This is unusual and worth flagging: the bot deploys massive volume in Soccer but loses money. Tennis is flat (+0.63%). NBA is where the directional model is sharpest (+9.62%). The esports bucket inside "Other" produces outsized per-market P/L (the best-market list includes multiple LoL and Valorant entries yielding +$8,000-$12,000 each on only 130-370 trades).

The price band distribution is broadly spread, with no single $0.10 bucket dominant:

The sub-$0.10 band is a massive outlier: 2.7% of trades, 0.9% of capital, 60.96% ROI, contributing +$83K - more than the total net P/L of the book. This is almost certainly hedge/longshot fills that occasionally hit at 10x+. The $0.10-$0.20 band bleeds (-10.76%), which is common in paired-cost books when one-side prices in the 15-18 cent range settle against the position.

The most important price observation is the $0.60-$0.80 zone: the bot deploys $5.2M there (35% of total capital) but generates -$57K combined. This is where the heavy favorite side of paired positions lives, and the problem is that $1.00 of expected payout is being purchased for more than $1.00 of combined outlay whenever the pairs cost above the break-even line.

The Order of Operations - One Event, Position by Position

Cavaliers vs. Pistons, NBA, 2026-05-07 (event: nba-cle-det-2026-05-07). Pistons resolved as the winner.

The bot entered this event across three related markets: the moneyline (Cavaliers vs. Pistons), a spread market (Spread: Pistons -6.5), and the moneyline again with varying conviction over roughly 11 minutes.

Walk-through:

1. The bot opens with moneyline Pistons at $0.77 - an 8-share position worth $6.20. This is the dominant leg: Pistons are the implied favorite at ~77 cents.

2. Within the same second, it enters Cavaliers at $0.24, spending $6.63 for 27 shares. The combined outlay is $12.83 for a paired position where one side must pay $1.00. Paired cost: $0.77 + $0.24 = $1.01 - just above break-even. The Cavaliers spend is actually larger in dollar terms here despite being the non-dominant side, because the cheap longshot buys many more shares per dollar.

3. Simultaneously it enters the spread market (Pistons -6.5) at $0.50/$0.50 near-even pricing, deploying $96.93 on Pistons cover and $109.26 on the Cavaliers side across several fills. The spread market is near-balanced, contributing close to $0 net edge.

4. The bot adds to the dominant moneyline position over the next 8 minutes (fills at $0.72-0.74), maintaining its asymmetric tilt on Pistons winning.

5. Pistons win. The dominant moneyline shares pay $1.00 each. The $6.63 Cavaliers hedge and $109 spread hedge both pay $0. Net: positive directional outcome; modest negative hedge tax.

The key structural observation: dominance ratio on this event's moneyline is approximately (Pistons USDC) / (Cavaliers USDC) = $19.44 / $20.00 = ~0.97x. That's near-parity on the moneyline! The bot's tilt shows up more clearly in the share count (44 Pistons shares vs 52 Cavaliers shares) and the price differential (paying $0.74-0.77 for Pistons vs $0.24-0.29 for Cavaliers). The Cavaliers position is the actual hedge; the Pistons is the real bet. This is standard both-sides construction where the cheaply-priced leg buys protection across more shares.

Why It Works - The Math

The strategy's core EV argument is that the directional model outperforms the market's implied probabilities in the high-conviction cases:

Dominance bucket analysis (3x+ markets):
  Markets:          4,505
  Dominant side WR:  73.0%
  Market-implied WR: ~62-68% (estimated from paired price distribution)

Edge per market (dom side):
  p_win_actual    = 0.730
  p_win_implied   = ~0.650 (typical mid-price in 3x+ dominant position)
  EV per $1 bet   = 0.730 * (1.00 / 0.650) - 1.00
                  = 0.730 * 1.538 - 1.00
                  = 1.123 - 1.00
                  = +$0.123 per dollar of dominant-side capital

Hedge tax per market (non-dominant side):
  Non-dominant wins 27% of the time, paying ~$3-4 per dollar at prices of $0.25-0.35
  Expected loss = 0.73 * (non-dom outlay) per market on average

Net EV: positive when directional edge exceeds hedge tax

The problem is the 1.0-1.5x bucket: 54.4% dom WR at near-50-cent prices is barely above break-even and the hedge tax nearly neutralizes the small directional edge. The P/L composition reflects this: the book's total spread P/L is -$288,341 (the spread mechanic is a net loss of $288K), which is partially offset by the directional wins. The entire +$96K net profit is generated by the 3x+ conviction tier. Everything below 3x is either marginally positive or negative.

This creates a clear strategic prescription: the bot should run narrower and higher-conviction rather than broad and paired. The current structure wastes capital on spread positions that erode value.

Phase 1 - Trader Profile

Scale and activity: - 892,903 BUYs, 0 SELLs, 27 active days (100% of calendar days) - $14,883,371 deployed across 10,243 unique markets - Average 33,070 trades/day; ~380 markets/day - 4,521 unique events, average 2.27 markets per event (confirming systematic multi-market coverage per event)

Trade size distribution:

The distribution is semi-power-law: median is small ($5.80) but there is a long right tail. The max at $2,040 is 352x the median, and the P99 is 26x the median. The top 5% of fills hold 39% of capital. This is consistent with a variable-clip sizer that scales clips with conviction, rather than a flat-clip bot.

The Lorenz data confirms: the bottom 50% of trades by size carry only 7.5% of capital, while the top 10% carry 52%. The sizing inequality is real and meaningful.

Execution signature: - Median inter-fill gap: 16 seconds - 40.2% of fills under 10s apart; 79.1% under 60s - Mean gap: 108 seconds (much higher, driven by long inter-event gaps) - Pattern: bot-speed bursts within a market, then slower inter-market transitions

The opening of the CSV shows characteristic fan-out: at 2026-04-11 00:00:41 UTC, six fills in the same second across two different Pistons vs. Hornets outcomes (422 shares at $0.50, 40 shares at $0.50, 19 shares at $0.50, 18 shares at $0.50, two more). Same-second multi-fill fan-out is a bot signature.

Active hours: Trading occurs all 24 UTC hours with no hard gap. Peak volume 14:00-20:00 UTC; secondary peak 11:00-14:00 UTC. The 03:00-07:00 UTC trough (13K-17K trades/hour vs 55K-68K at peak) is not zero - the bot is running overnight, just at reduced intensity when fewer live events are available.

Archetype: BOTH-SIDES MM + DIRECTIONAL OVERLAY - 79.1% both-sides rate with a strong conviction curve. Not a pure MM (paired cost is underwater) and not a pure directional (hedge legs are always present). A hybrid that extracts value primarily from directional accuracy.

Phase 2 - Core Strategy Identification

Both-sides participation rate: 79.1% (8,103 of 10,243 markets)

This is a both-sides book. For 8 out of every 10 markets the bot touches, it buys both YES and NO (or both team outcomes). This is the defining structural characteristic.

However, the paired cost tells us the MM mechanism is not the profit source: - Median paired cost: $1.026 - Mean paired cost: $1.033 - % of paired markets with cost below $1.00: 41.0% - % with cost below $0.97: 30.5%

At mean paired cost of $1.033, every matched-share pair costs $0.033 more than it pays out on average. The spread engine is a net negative. This collapses the pure-MM hypothesis.

What IS the profit source: The dominance ratio analysis (Phase 3). When the bot tilts 3x+, it wins 73% on the dominant side. The P/L is directional.

Archetype classification: - NOT pure Market Making (paired cost underwater, spread P/L = -$288K) - NOT pure Directional (79.1% of markets have a hedge leg) - IS a directional model expressed through asymmetric pairing: the "hedge" leg is insurance that costs ~3.3 cents per dollar of matched notional, while the "dominant" leg captures the directional edge when the model is correct

The strategy is closest to what quantitative sports modelers call a "rounded book" approach: you have a price for every outcome, you always take positions on all sides, but you size them proportional to your edge, making the dominant-side positions larger whenever your model's edge is highest.

Phase 3 - Dominance Ratio Analysis

This is the most important phase for this wallet.

Four observations from this table:

1. The win rate jump from 2-3x to 3x+ is the largest discrete step in the conviction curve - from 61.1% to 73.0%. The model's edge is not linear; it accelerates at the top.

2. Paired cost declines as conviction rises - from $1.054 in the 1-1.5x bucket to $1.020 in the 3x+ bucket. When the model is confident, the bot buys the dominant side at a better relative price (because the "cheap" hedge is at a lower price, reducing the blended cost). This is a favorable structure.

3. The 3x+ bucket contains 4,505 markets - 55% of all both-sides markets - meaning most of the book is in the high-conviction zone. The model fires at high confidence more often than not.

4. The 1.0-1.5x bucket (1,348 markets, 54.4% WR, $1.054 paired cost) is the value-destroying segment. These are markets where the bot has nearly no edge and is paying over-round for the privilege of covering both sides. Eliminating this segment would likely improve total P/L.

The practical implication for replication: only enter when dominance is 3x or higher on the dominant side. The 1-2x zone is noise.

Phase 4 - Entry Price Analysis

The per-band breakdown (Phase 1 table above) reveals three regimes:

Regime 1: Sub-$0.10 (longshot/hedge fills) - 60.96% ROI on $136K. These are the extreme-underdog legs of paired positions: when the model puts all the weight on the favorite side and buys the opponent at $0.03-$0.09 as pure insurance. The few that win pay 10-30x. The ROI is spectacular but the capital is tiny.

Regime 2: $0.40-$0.60 (coin-flip zone) - +1.75-2.83% ROI, the spread-capture sweet spot. The bot deploys heavily here ($1.27M + $2.50M = $3.77M in the two buckets). These are the "balanced" sides of paired positions where the spread mechanic comes closest to break-even.

Regime 3: $0.60-$0.80 (heavy-favorite zone) - Both bands are negative ROI (-0.76% and -1.45%). The bot deploys $5.2M at $0.60-$0.80 and loses $57K. This is the dominant-side leg of paired positions where the model's directional call is expressed, but in markets where the paired cost exceeds $1.00 by enough to overwhelm the directional win rate.

Sub-bucket inspection: The price distribution across the full 100-cent range shows no single dominant tick. This is unlike a pure latency-arb bot (which concentrates at specific ticks) or a penny-floor bot. The entry price is wherever the orderbook offers value when the model fires.

The $0.10-$0.20 bucket (-10.76% ROI) is notable: 36,000 trades, -$36K P/L. These are probably the short-odds hedge legs of high-conviction dominant positions - the bot pays $0.15 for the underdog side when it has 6:1 conviction on the other side. The underdog rarely wins, and the ROI is a consistent drain.

Phase 5 - Category and Vertical Breakdown

Soccer is the single largest source of losses (-$58,634) despite being the second-largest category by volume. The bot's model performs poorly on Soccer relative to tennis and American sports. This may reflect the difficulty of modeling soccer outcomes (low-scoring, high variance) versus more deterministic markets (basketball final scores).

The esports component within "Other" is a highlight: multiple LoL markets appear in the best-performing list with +$8,000-$12,000 per market. The LoL: Dplus KIA vs KT Rolster, LoL: KT Rolster vs HANJIN BRION, and LoL: LOUD vs Vivo Keyd Stars markets all yielded substantial positive P/L on moderate trade counts. The model appears to have real edge in structured esports competitions.

Top performing individual markets: - Trail Blazers vs. Spurs: +$71,421 on $21K volume - anomaly, likely a massive underprice that resolved favorably - LoL: Dplus KIA vs KT Rolster (G2): +$12,971 - Will Portsmouth FC win: +$12,454 on $2.6K (extremely high ROI on a tiny Soccer bet) - LoL: KT Rolster vs HANJIN BRION (BO3): +$10,197

Worst markets: - Knicks vs. Hawks: -$12,435 on $35K volume - LoL: Weibo Gaming vs Anyone's Legend (BO3 + G2): -$9,454 + -$8,124 (both 0 wins - model was completely wrong on this matchup) - Madrid Open: Tsitsipas vs Ruud: -$7,424 - Lakers vs. Rockets: -$7,127

Phase 6 - Timing and Execution

Hourly P/L pattern (UTC):

The hourly pattern is counterintuitive: the highest-volume hours (14:00-19:00 UTC) are the worst-performing, while several low-volume overnight hours produce the best P/L. This suggests the model's edge is highest in low-competition, off-peak windows where fewer efficient actors are participating. The 09:00 UTC hour (+$47K, 60.3% WR) is the exception - European morning, high-quality Tennis and Soccer events, where the model has strong signal.

Day-of-week:

Saturday has 2.4x the weekday average trade count (primarily Tennis and Soccer weekend schedules) but loses money. The model's edge on weekend sports is lower than on weekday events. Wednesday is the best day (+$69,703) with the highest win rate (57.2%) - mid-week NBA, Tennis, and soccer with strong model performance.

Second-side lag: Median 634 seconds (10.6 minutes) between entering first and second side in a paired market. This is consistent deliberate pairing, not coincidental. The bot has a defined time window for hedging after the dominant-leg entry.

Phase 7 - Filter Experiments

The high-conviction filter is the most transformative finding in this entire analysis. Keeping only the dominant side of markets where dominance ratio is 2x or higher - 376,315 trades, 42% of the book by count but 46% by capital - yields +$1.258M P/L at +18.33% ROI. That is roughly 13x the actual net P/L on 46% of the capital. The remainder of the book (the hedge legs and low-conviction paired positions) destroys approximately $1.16M of value.

This tells us directly: the bot is subsidizing its losing book segments with its winning directional calls. A stripped-down version running only high-conviction dominant sides would dramatically outperform the current structure.

The "exclude worst hours" filter performs poorly in absolute terms - removing hours 01:00, 02:00, 04:00, and 07:00 (the identified worst hours by win rate) actually decreases absolute P/L, because hours 02:00 and 03:00 have low win rates but large positive P/L from high-conviction esports and soccer positions. Do not filter by hour.

Phase 8 - Rolling Window Consistency

The 15-day rolling windows are nearly all green (96.3%), demonstrating that the edge is real and persistent over multi-week periods. The 7-day windows are more volatile - 66.7% green - with several large negative 7-day windows in late April/early May. The rolling7 data shows a sharp drawdown sequence: -$31K (Apr 29), -$75K (Apr 30), -$81K (May 1), -$20K (May 2) before recovery. This corresponds to a week when Soccer (perennially the weakest category) dominated the event schedule.

The edge is real but not monotonic. The strategy has a structurally positive 15-day expectation but meaningful 7-day drawdown risk from category concentration (when a bad Soccer week dominates the book) and model variance on high-volume events.

Phase 9 - P/L Decomposition

The decomposition reveals the true anatomy of the P/L. The hedge tax is enormous: $5.3M spent on the non-dominant legs of paired positions, with a win rate on those legs of roughly 27% (100% - 73% where dominant side wins). The directional wins on the dominant side must overcome both the raw hedge cost and the spread over-round.

Directional P/L (estimated):

High-conviction (3x+) dominant leg capital:  ~$6M (estimated from 376K trades × avg $16)
High-conviction win rate:                      73%
Expected payout:  0.73 × ($6M / avg_price)   >> positive
Hedge tax on these markets:                   ~$1.5-2M (non-dominant legs)
Net directional gain (estimated):             ~$1.3M before hedge tax
After hedge tax:                              ~$400-500K
Plus low-conviction directional gains:        ~-$200K (drag)
Less spread over-round:                       -$288K
Net:                                          ~$96K (matches actual)

The conclusion is consistent: the entire net P/L is generated by the 3x+ conviction tier, and the rest of the structure is a wash at best, a drag at worst.

Phase 10 - Strategy Specification (short form)

One-sentence summary: A both-sides prediction market bot with a calibrated directional model that expresses conviction through asymmetric allocation within paired positions, generating all net P/L from markets where the dominant side is tilted 3x or higher.

Edge source: Directional model accuracy in the high-conviction tier (73% dominant win rate at 3x+). The spread mechanic is structurally negative and provides only risk-reduction, not alpha.

What works: NBA markets (+9.6% ROI), esports within Other (+strong individual market returns), NHL (+3.3%), the sub-$0.10 longshot wing (+60.9% ROI), Wednesday scheduling, 09:00 UTC hour.

What drags: Soccer (-1.6% ROI, -$58K absolute), the 1.0-1.5x dominance bucket (54.4% WR, paired cost $1.054), the $0.60-$0.80 price band (-$57K), weekend schedule bias.

What replicators must NOT do: run both-sides positions in the 1-2x dominance bucket. The hedge cost exceeds the directional edge at that conviction level. See playbook for the stripped-down spec.

Wallet: 0x84ad9c5c547a82ec9a08547b94bd922446e5bfb7 Window: 2026-04-11 → 2026-05-07 (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: 79.11% (8,103 of 10,243 markets)
• Median paired cost: $1.0256
• Mean paired cost: $1.0329
• Paired cost % under $1.00: 41.0%
• Paired cost % under $0.97: 30.5%
• Median 2nd-side hedge lag: 634s

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: 19 of 27 (70.4%)
• Rolling 7-day P/L range: -$81,058 → +$106,170
• Rolling 15-day windows green: 26 of 27 (96.3%)
• Rolling 15-day P/L range: -$10,463 → +$132,795

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 2026-05-08 08:59 UTC.

Wallet: 0x84ad9c5c547a82ec9a08547b94bd922446e5bfb7 Window: 2026-04-11 → 2026-05-07 (27 days) Baseline: 892,815 resolved BUYs · 55.67% WR · $14.87M deployed · +$96,986 P/L · +0.65% ROI

The headline result

One filter produces transformative lift. Most others are destructive or negligible. The high-conviction dominant-side filter increases realized P/L from +$97K to +$1.258M - a 13x improvement - by removing the hedge legs and low-conviction pairings that drag the book. Every other filter produces modest at best, often harmful, results.

The price-band filter does minor good (+0.24% ROI improvement) but cuts 46% of absolute P/L. The category filter on NBA surfaces the strongest per-dollar performance but captures only 2.8% of trades. The hour filter is outright destructive. The stacked combined filter produces a reasonable ROI% on a negligible absolute base.

Filter results table

Filter-by-filter commentary

1. Price band filter ($0.30-$0.70) → MARGINAL, USE WITH CAUTION

Applying the standard sweet-spot price filter reduces the book from 892K to 482K trades and shifts ROI from +0.65% to +0.89% - a real but modest improvement of 24 basis points. Absolute P/L drops from +$97K to +$68K, a loss of $29K in net dollars.

The mechanism: the filter removes the extreme ends of the distribution. The sub-$0.10 band (+60.96% ROI, +$83K) is stripped out because those fills are below $0.30, representing a major loss of a high-ROI segment. The $0.70-$0.90 band (-0.76% to +0.06% ROI) is also stripped, removing a modest drain. The net effect is marginally positive in percentage terms but negative in absolute dollars because the sub-$0.10 longshot wing is so ROI-dense.

Verdict: Do not apply this filter if replicating for absolute dollar returns. Apply it only if capital is limited and you need to concentrate on the most ROI-efficient tier of the $0.40-$0.60 price zone.

2. High-conviction filter (dom 2x+, dominant leg only) → TRANSFORMATIVE LIFT

This is the most important filter verdict in the entire report. Keeping only the dominant side of markets where the bot's allocation ratio is 2x or higher - and discarding all hedge/non-dominant legs - produces 376,315 trades, $6.87M capital, +$1.258M P/L, +18.33% ROI.

Compare to the unfiltered baseline: +$97K on $14.87M (+0.65%). The high-conviction dominant-leg subset alone produces 13x more absolute profit on 46% of the capital, for an ROI that is 28x higher.

The math is compelling:

High-conviction dominant legs (2x+):   +$1,258,243 P/L  on $6.87M  (+18.33%)
Everything else in the book:           -$1,161,257 P/L  on $8.00M  (-14.52%)
Total book:                            +$96,986 P/L     on $14.87M (+0.65%)

The "everything else" - hedge legs, non-dominant legs, and low-conviction paired positions - destroys $1.16M of value that the directional winners barely offset. This is the structural cost of the current both-sides architecture.

Practical implication for replication: A bot that only trades the dominant side in 2x+ conviction markets, ignoring all hedging, would generate 13x the current P/L on less capital. The hedge discipline adds robustness and reduces variance but costs a massive amount in expected value.

3. Category filter (NBA only) → MEANINGFUL_LIFT in ROI, SMALL absolute book

Isolating NBA trades yields 24,777 trades, $1.06M deployed, +$101,762 P/L, +9.62% ROI. This is the single best-performing category by ROI margin and narrowly beats the book's total net P/L in absolute terms on only 7.1% of capital.

The NBA filter captures the sharpest directional signal in the book. The model's win rate on NBA dominant positions is meaningfully higher than on Soccer or Tennis. The trade-off is that the NBA schedule is narrow - roughly 82 regular-season games plus playoffs, limiting daily coverage to 0-6 games depending on the schedule. The per-market P/L is high (best-market table shows NBA entries consistently in the -$2K to +$2K range per 500-750 trade market) but the total capacity is capped.

Esports within "Other" shows similar characteristics (high per-market ROI on individual events) but is less cleanly parseable from the category breakdown.

Verdict: Over-weight NBA and esports relative to Soccer. The Soccer allocation is the single biggest drag on total P/L.

4. Soccer exclusion → MEANINGFUL_LIFT

While not listed as a named filter in the output, Soccer exclusion is implied by the category analysis. Soccer has -$58,634 P/L on $3.75M deployed (-1.56% ROI). Removing Soccer from the book while keeping everything else would increase total P/L by $58,634, from +$97K to +$155K, and increase overall ROI from 0.65% to 1.39% on the remaining $11.1M.

The bot deploys enormous Soccer volume (366K trades, 41% of total trade count) for negative returns. The directional model is weakest in Soccer, likely because soccer outcomes are inherently higher-variance (low-scoring, draws possible, strong upset frequency) than basketball or tennis.

Verdict: If you can replicate the model's architecture without Soccer coverage, do so. If you must trade Soccer, concentrate it in the 3x+ conviction tier only and monitor weekly.

5. Hour exclusion filter → DESTRUCTIVE - DO NOT APPLY

The "worst 4 hours" filter (hours 01, 02, 04, 07 UTC identified as having lowest win rates at 53.4-52.6%) removes $1.52M of capital and drops absolute P/L by $56K, from +$97K to +$40K. This is counterintuitive: the hours with the lowest win rates include hours 02:00 (+$57,185 P/L) and 03:00 (+$23,137 P/L).

The explanation: win rate is an imperfect proxy for P/L in a both-sides book. Overnight hours have low win rates because the bot is placing mostly hedge legs (cheap non-dominant sides) during low-event periods. But some of these hedge positions pay very large sums when they win (longshot $0.05 positions paying $1.00 each when the upset lands). Hour 02:00 UTC is the second-most profitable hour in the book despite having the second-lowest win rate.

Verdict: Do not filter by hour on this wallet. The win rate vs P/L relationship is non-monotonic due to the asymmetric payoff structure of longshot hedge positions.

6. Stacked combined filter → HIGH ROI%, TINY CAPACITY

The precomputed combined filter (high-conviction + top category + $0.30-$0.70) yields 8,603 trades, $346K capital, +$20,585 P/L, +5.95% ROI. The ROI% is attractive (5.95% vs 0.65% baseline) but the absolute capital deployed drops by 98%, from $14.87M to $346K. This is essentially the intersection of NBA + high-conviction + coin-flip prices, which is a tiny subset of the total book.

The stacked filter demonstrates diminishing marginal intersection but confirms the direction: tighter, higher-conviction, NBA-focused filters improve ROI%. The trade-off is capacity.

What filters WOULD add value if they were measurable

The standard filter battery misses the most actionable dimension of this trader. The genuine improvement filters would require data beyond the trade CSV:

Bottom line

Three concrete recommendations for a replicator:

1. Apply the high-conviction dominant-side filter ruthlessly. Only execute the larger side when your allocation ratio is 2x or higher. The sub-2x positions contribute negative absolute value after accounting for hedge costs.

2. Remove Soccer from the universe, or at minimum restrict Soccer to 3x+ conviction only. The Soccer allocation is the single largest drag, responsible for -$58K P/L on 25% of total capital.

3. Do not filter by hour. The win rate by hour is misleading because overnight hours contain high-value longshot positions. Filter by conviction ratio, not by clock.

The single most useful action for this strategy is not adding a filter - it is removing the structural incentive to hedge low-conviction positions. See the playbook for the implementation spec.

Source wallet: 0x84ad9c5c547a82ec9a08547b94bd922446e5bfb7 Strategy: Both-sides directional overlay - paired positions with asymmetric tilts driven by a sports model Reference book: $14.87M BUY notional → +$96,765 net P/L → +0.65% ROI in 27 days Stripped high-conviction version: $6.87M → +$1.258M → +18.33% ROI (dominant side only, 2x+ conviction)

One-paragraph operator brief

Build a Polymarket bot that maintains a calibrated directional model for sports and esports outcomes (basketball, hockey, tennis, esports at minimum). For each live event, compute your fair-value probability for each outcome and compare against the CLOB mid-prices. When your model's dominant-side edge implies an allocation ratio of 2x or greater relative to the hedge side, execute the dominant-side position at market. Optionally execute the hedge side as well (for variance reduction), but only size it at a ratio proportional to your confidence. Cap Soccer exposure. Over-weight NBA, NHL, and esports. Run 24/7 but expect better model performance during European morning (08:00-11:00 UTC) and mid-week. Expect +18% monthly ROI on capital deployed in the high-conviction dominant tier; expect the hedge tax to erode returns if you systematically over-hedge.

1. Market selection

Category priority order (by ROI performance):

1. NBA: +9.62% ROI - highest priority, size aggressively when high-conviction
2. Esports (LoL/CS/Valorant): strong individual market returns, prioritize BO3/BO5 formats
3. NHL: +3.27% ROI - solid, proceed
4. NFL/MLB: +1.41% / -0.73% - moderate, use 3x+ conviction threshold strictly
5. Tennis: +0.63% - deploy but monitor; volume is large so even thin edge compounds
6. Soccer: -1.56% - skip moneylines unless 3x+ conviction, restrict to major leagues

Multi-market coverage per event: The reference wallet covers multiple markets per event (moneyline, spread, totals). This is recommended for capital efficiency - the model signal applies across all derivative markets of the same underlying event. Build slug parsing to identify event clusters (e.g., all markets sharing nba-cle-det-2026-05-07 as event slug).

2. Entry logic

The entry signal is your model's output versus the CLOB price. The key threshold is the dominance ratio that the reference wallet exhibits when it wins at high rates.

def should_enter(market, outcome, model_prob, clob_mid):
    # Category gate
    if market.category == "Soccer" and model_prob < 0.65:
        return None  # avoid Soccer below strong conviction

    # Compute edge
    edge = model_prob - clob_mid

    # Minimum edge threshold
    if abs(edge) < 0.04:
        return None  # less than 4% edge is noise

    # Dominance gate: only enter if your allocation to this side 
    # vs the hedge side is 2x or higher
    hedge_prob = 1.0 - model_prob
    implied_dom_ratio = model_prob / hedge_prob
    if implied_dom_ratio < 2.0:
        return None  # low-conviction zone, reference book shows <58% WR here

    direction = "dominant" if edge > 0 else None
    return direction

Do NOT enter the following: - Sub-$0.10 speculative buys on unknown underdogs without a model backing - Both-sides positions where your dominance ratio is below 2x - Soccer markets where your model confidence is below 60% - Any market in the final 60 seconds before resolution (noise trading zone)

3. Exit logic

The reference wallet holds all positions to resolution - zero SELLs across 892,903 trades and $14.87M deployed. This is the intended exit mechanism: position resolves at $1.00 (win) or $0.00 (loss).

There is no SELL engine to replicate. Positions are held from entry through settlement. The only "exit" decision is whether to add to a position if the price moves favorably after entry (partial averaging). The reference CSV shows the bot adding to positions over a 10-minute accumulation window when prices shift (e.g., multiple Pistons fills at $0.71, $0.72, $0.74, $0.77 at different timestamps).

def manage_position(position):
    # Hold to settlement - no active exit
    # Optional: add to position if price improves
    current_mid = get_clob_mid(position.market, position.outcome)
    if current_mid < position.avg_entry_price * 0.95:
        # Price improved 5% since entry - add at better price
        add_clip = min(remaining_budget, base_clip * 0.5)
        enter(position.market, position.outcome, add_clip)
    # Otherwise: do nothing, let it settle

Settlement timing: Markets resolve at event completion. Basketball games typically resolve within 3-4 hours of tip-off; Tennis matches within 2-4 hours; Esports BO3 matches within 3-5 hours. Capital cycles within half a day for most market types.

4. Sizing model

The reference wallet uses variable clip sizing correlated with conviction. The sizing profile:

Bankroll scaling:

The reference wallet deploys an average of $551K/day across all markets but this includes many low-conviction and hedge positions. The pure high-conviction dominant tier deploys approximately $6.87M over 27 days, or ~$255K/day. Scale your bankroll expectations accordingly.

No Martingale, no doubling after losses. The reference wallet shows no evidence of loss-chasing scaling. When a high-conviction position loses, the next position is sized identically by model output, not by the outcome of the prior bet.

5. Both-sides allocation

The reference wallet hedges 79.1% of its markets. Based on the filter analysis, systematic hedging at low conviction ratios destroys $1.16M of value. The recommended approach for a replicator:

Simplified: Dominant side only (recommended for replicators) - Enter only the dominant side (2x+ conviction) - Never enter the non-dominant side as a hedge - Expected ROI: ~+18% on deployed capital (based on filter result) - Variance: higher per-market, but 73% WR at 3x+ is sufficient

Advanced: Paired with selective hedging (replicating the reference structure) - Enter dominant side at full clip when dom ratio >= 3x - Enter hedge side at 20-30% of dominant notional when dom ratio is 3-5x - Enter hedge side at 10-15% of dominant notional when dom ratio is 5x+ - Skip hedge entirely when dom ratio is below 3x (reference wallet's 2-3x tier loses money overall)

The second-side lag in the reference book is a median of 634 seconds (10.6 minutes). If you do hedge, execute the hedge leg within 10-15 minutes of the dominant entry to match the reference's timing discipline.

6. Sports model requirements

The edge in this strategy is 100% model-driven. Without a calibrated sports model, you cannot generate the 3x+ conviction signals that produce the 73% win rate. The model minimum viable requirements:

For NBA: - Team strength ratings (Elo or similar) - Home/away adjustment - Rest days, back-to-back game flag - Injury reports (starter availability) - Recent form (last 5-10 games)

For Tennis: - Surface-adjusted Elo ratings (hard / clay / grass separate) - Head-to-head record - Recent match outcomes (last 4-8 weeks) - Tournament round (fatigue in later rounds)

For Esports (LoL/CS2): - Team rating within region (LCK, LEC, LPL, etc.) - Match format (BO3 vs BO5 - higher-rated teams cover BO3 at higher rates) - Tournament stage (group vs bracket) - Recent match history (last 3-5 matches) - Roster stability flag

For Soccer: - Use extreme caution. The reference book loses -$58,634 in Soccer. A basic Elo model is insufficient. If including Soccer, you need: expected goals data, recent xG performance, possession stats, and squad rotation likelihood. Only trade 3x+ conviction Soccer.

Model output: A probability estimate for each outcome. Comparing this against the Polymarket CLOB mid-price gives you the edge. The bet size scales with the edge magnitude.

7. Hour scheduling and operational requirements

DO NOT apply a hard off-window. The reference wallet trades all 24 hours and several overnight hours produce outsized P/L (hour 02:00 UTC = +$57K, hour 23:00 = +$33K). The overnight hours that appear weak by win rate are profitable by absolute P/L.

Infrastructure requirements:

8. Risk profile and sizing guardrails

Hard position limits:

Max per single market:        $1,500 (at $50K bankroll)
Max per single event cluster: $3,000 (across all derivative markets)
Max Soccer exposure:          20% of daily deployment budget
Daily drawdown stop:          -3% of NAV
Rolling 7-day drawdown stop:  -8% of NAV → pause and audit model

9. Diagnostic checklist - "is the bot still working?"

Run weekly:

10. What this playbook deliberately does NOT include

• No hedge-first architecture. The reference book's spread mechanic loses $288K. Building a system that tries to capture spread income is the wrong frame; the alpha is directional.

• No fixed paired-cost target. Chasing a paired cost below $1.00 (the classic MM pitch) would require buying very thin books at unfavorable prices on the non-dominant side. The reference wallet already shows what happens at $1.033 mean paired cost: the spread leg destroys value. Don't try to fix this by tightening the paired cost; fix it by not pairing at low conviction.

• No hour-based scheduling constraints. The filter analysis shows hour exclusions are destructive. Run 24/7 with uniform entry logic; the market schedule determines natural volume fluctuation.

• No Soccer-heavy deployment. Soccer is 41% of the reference wallet's trade count and the only major negative-ROI category at -1.56%. A replicator does not need to match the reference wallet's Soccer volume. Concentrating on NBA, Tennis, NHL, and Esports with strict Soccer gates will outperform the reference on a per-dollar basis.

• No longshot speculation without model backing. The sub-$0.10 band has +61% ROI in the reference book but those positions are organic outputs of the paired-position structure (cheap hedge sides). Buying sub-$0.10 positions speculatively without the model foundation will lose money.

• No copy-trading of this wallet's specific picks. Sports markets close after each event. By the time another wallet's fill appears in the on-chain data, the match has likely progressed or resolved. This strategy cannot be copy-traded; it requires an independent model.

The core discipline: high conviction, dominant side, American sports and esports as primary verticals. The reference wallet's broad Soccer coverage is a legacy of the model's application to all available markets, not a deliberate alpha source. Run narrower, run sharper, collect the directional wins.

TL;DR - implementable in ~200 lines of Python

# Pseudocode - model is your own implementation. 
# This shows the execution loop only.

async def run_bot():
    model = SportsModel()       # your calibrated prob model
    clob = PolymarketCLOB()

    while True:
        live_markets = await clob.get_active_markets()

        for market in live_markets:
            # Category gate
            if market.category == "Soccer" and not is_high_quality_soccer(market):
                continue

            # Get model estimate and CLOB mid
            for outcome in market.outcomes:
                model_prob = model.predict(market, outcome)
                clob_mid   = await clob.get_mid(market, outcome)
                edge       = model_prob - clob_mid

                if abs(edge) < 0.04:
                    continue

                # Dominance gate
                dom_ratio  = model_prob / (1.0 - model_prob)
                if dom_ratio < 2.0:
                    continue

                # Sizing
                clip = compute_clip(dom_ratio, available_capital())
                # dom_ratio 3x+: full clip; 2-3x: 60% clip
                if dom_ratio >= 3.0:
                    clip = clip * 1.0
                else:
                    clip = clip * 0.6

                clip = min(clip, MAX_PER_MARKET)

                # Execute dominant side
                await clob.buy(market, outcome, clip)

                # Optional hedge (skip if capital-constrained)
                if ENABLE_HEDGING and dom_ratio >= 3.0:
                    hedge_outcome = market.other_outcome(outcome)
                    hedge_clip    = clip * 0.20  # 20% of dominant notional
                    await clob.buy(market, hedge_outcome, hedge_clip)

        await asyncio.sleep(15)  # poll every 15 seconds

Deploy with $15K-$50K working capital. Expect +15-20% monthly ROI on capital deployed in the high-conviction tier. Monitor Soccer weekly. The model is the product; the execution is the plumbing.