FerrariChampions2026

Wallet: 0xfe787d2da716d60e8acff57fb87eb13cd4d10319 Window: 2026-04-20 to 2026-05-19 (30 calendar days, 30 active) Universe: 233,517 trades · 7,628 markets · 3,726 events · $43,768,856 gross turnover Net P/L: +$1,136,314 on $43,554,480 resolved BUY notional = +2.60% ROI in 30 days

The Punchline

This is a both-sides spread-capture engine running across sports prediction markets at industrial scale. The wallet buys YES and NO on the same game, locking in the bid-ask spread as guaranteed profit at settlement, then skews capital allocation toward the side it believes will win. It is not a pure market maker (the skews are large and informative) and not a pure directional bettor (it hedges virtually every position). It is a hybrid: systematic spread capture with a directional model layered on top.

The economics require understanding the decomposition. Spread P/L is +$1,776,733 - the aggregate guaranteed profit from all paired positions where combined cost was below $1.00. That number is eroded by the hedge tax (-$16,553,248), the capital spent on the losing side of every matched pair. The net of those two streams, plus directional wins on unmatched excess, produces the realized +$1,136,314. The spread engine is working; the hedge tax is simply the structural cost of guaranteeing it.

The dominance-ratio analysis confirms a genuine directional signal embedded in the sizing. Dominant-side win rate escalates monotonically from 49.7% at 1.0-1.5x tilt to 59.6% at 3.0x+. This is not random: the bot's model is predicting winners at above-chance rates when it tilts hard, adding incremental alpha on top of the structural spread. The NBA vertical is the one place this model failed systematically, generating -$191,047 of losses and partially offsetting large wins in tennis and MLB.

---

What He Trades

The universe is broad but structured across three primary categories:

Tennis is the workhorse. The 33.9% raw win rate on tennis BUYs looks alarming until you understand the spread structure: a pure spread-capture bot buying both YES and NO will have an overall win rate close to 50%, but the individual side that lost the match shows 0% WR on those tickets. The tennis book's 3.41% ROI on $17.2M is the spread engine performing as designed.

The "Other" category at +8.87% ROI is the highest-ROI meaningful bucket. Examining the best-markets table, this includes UFC/MMA fights and some over/under totals that are not sport-specific. The UFC/MMA standalone at +35% ROI on $86K is a statistical outlier given the sample size of 230 trades.

NBA is the failure mode. The top-10 worst markets list is dominated by NBA series: Timberwolves vs. Nuggets (-$84K), Rockets vs. Lakers (-$83K), Cavaliers vs. Raptors (-$81K), Timberwolves vs. Nuggets (second instance, -$73K), Nuggets spread (-$70K), Spurs vs. Timberwolves (-$69K), Timberwolves vs. Spurs (-$65K), Hawks vs. Knicks (-$63K). Total identifiable NBA losses exceed -$550K across those eight markets, partially offset by wins like Rockets vs. Lakers (+$60K in a separate instance) and 76ers vs. Celtics (+$25K).

---

The Order of Operations - One Market, Trade by Trade

The clearest illustration of the strategy is Magic vs. Pistons (nba-orl-det-2026-04-19), where the bot entered at the very start of the window (April 20, 00:00-00:33 UTC) and built a large paired position over 33 minutes. The game had already been played (it was April 19 Eastern time); the market was still open for resolution.

The bot's behavior here shows the full playbook:

Walk-through:

Phase 1-4: The bot identifies Magic vs. Pistons as a live market and begins buying Magic (the outcome that ultimately resolves as the winner) at $0.36-0.51. These are the "dominant side" fills - the bot's model favors Magic, so it builds a large Magic position first.

Phase 5-6: Approximately 30 seconds into the Magic accumulation, the first Pistons hedge fills appear at $0.47. At this point the implied paired cost is roughly $0.51 + $0.47 = $0.98, which is below $1.00 - the spread is locked.

Phase 7: The bot purchases more Pistons at $0.66-0.68 - after the price has moved significantly. This is the hedge-tax regime: it's paying more for the hedge than the original signal price, which increases the paired cost. This is the second-side lag consequence: prices move in the 14-minute median gap between first and second side.

Phases 8-10: Late Pistons adds at $0.36-0.43 as the market drifts toward the winning outcome. The bot is still hedging with a remaining Pistons exposure.

The resolution: Magic wins, as the Resolved Side column confirms for all Magic tickets. All Pistons tickets returned $0. The spread on this market was a large net win for the Magic side, offset by the Pistons hedge cost.

---

Why It Works - The Math

The spread-capture mechanism is straightforward:

Paired cost = YES_VWAP + NO_VWAP per paired unit

If paired_cost < 1.00:
  guaranteed_profit_per_unit = 1.00 - paired_cost
  (resolution pays $1.00 on the winning side, $0.00 on losing side)

Median paired cost observed: $0.9748
Median spread per unit: $0.0252

At scale:
  Markets with both-sides: 4,202
  Average notional per paired market: ~$10,400
  Spread P/L = 1,776,733 (confirmed by decomp)

The directional add-on:

Dominance ratio → Dominant-side win rate:
  1.0-1.5x:  49.7% wins (685 markets) - pure spread, no signal
  1.5-2.0x:  52.1% wins (454 markets) - weak signal
  2.0-3.0x:  58.8% wins (566 markets) - real signal
  3.0x+:     59.6% wins (2,497 markets) - strong signal

At 3.0x+ dominance:
  Win rate 59.6% vs. 50% baseline = +9.6% hit-rate alpha
  Applied to 2,497 markets, this is a meaningful incremental P/L contribution

The strategy's vulnerability is two-fold. First, the hedge tax is enormous: -$16.5M of capital spent on losing sides, exceeding the gross spread P/L of +$1.8M. The net is positive, but the margin for error is thin - a sustained model failure in a high-volume category (as NBA demonstrates) can erase multiple weeks of spread profit in days. Second, the second-side lag creates paired costs above $1.00 on a meaningful fraction of positions (the 40.2% of both-sides markets where paired cost exceeded $1.00).

---

Phase 1 - Trader Profile

Scale and activity:

• 233,517 BUYs, 0 SELLs across 30 days
• $43,768,856 total BUY notional
• 7,784 trades per day average
• 7,628 unique markets, 3,726 unique events
• 30 of 30 calendar days active

Trade size distribution (extreme power-law):

The 24× difference between median ($6.60) and mean ($187.43), and the top-5% concentration at 71.6%, confirm a power-law distribution. The bot fires many small probe fills and occasional enormous conviction clips. The single $190,399 fill is a one-off extreme event; the P99 of $3,335 is more representative of the large-clip regime.

Execution signature:

• Median inter-fill gap: 5.0 seconds
• 57.4% of consecutive fills under 10 seconds
• 74.1% under 60 seconds
• Mean gap: 199.4 seconds (distorted by overnight pauses)

The 5-second median with 57.4% under 10 seconds is bot-speed execution. Same-second multi-fill fan-outs are visible throughout the CSV (e.g., multiple fills on mlb-bos-kc-2026-05-19-spread-home-2pt5 at 23:58:59 and 23:59:01 UTC).

---

Phase 2 - Core Strategy Identification

Both-sides participation rate: 55.1% (4,202 of 7,628 markets had both sides purchased)

This is above the 60% threshold for "probably MM" from the spec, but the rate combined with active directional tilting makes this a hybrid. Classification:

Primary: A (Both-Sides Spread Capture / Market Making)

Evidence:

• 55.1% both-sides rate
• Median paired cost $0.9748 (below $1.00 - guaranteed spread)
• 59.8% of both-sides markets have paired cost below $1.00
• Spread P/L = +$1,776,733 confirms the mechanism

Secondary: B (Directional Betting) layered on top

Evidence:

• Dominant-side win rate escalates from 49.7% at 1-1.5x to 59.6% at 3x+
• Second-side lag of 836s (not simultaneous pair - directional entry first, hedge second)
• Single-side markets (44.9% of book) are pure directional calls
• Top-performing markets include significant one-sided plays (e.g., Athletics vs. Angels: +$121,624 with no counterpart hedge detectable)

The bot is NOT:

• A pure market maker (44.9% of markets are one-sided; dominant-side win rates exceed 55% at high tilt)
• A copy-trader (execution speed rules this out)
• A stale-price sniper (entries span all price bands, not concentrated at near-certainties)

---

Phase 3 - Dominance Ratio Analysis

The monotonic escalation in dominant-side win rate from 49.7% to 59.6% as dominance increases from 1x to 3x+ is the critical finding. This is not random variation across 4,202 markets - it is a genuine directional signal embedded in the sizing scale.

The mean paired cost at 3.0x+ (0.9829) is actually slightly higher than at 2.0-3.0x (0.9505). This reflects the high-dominance regime: when the bot is extremely confident in one side, it may be bidding higher on the spread, or more likely the dominant side's price is already elevated, making it harder to get a tight paired cost. The 3x+ bucket is the "I think I know who wins" mode where the directional add-on is most valuable but the spread is thinner.

The 1.0-1.5x bucket at exactly 49.7% dominant-side win rate confirms that near-even splits are pure spread plays with no directional information. The bot correctly uses this tier for guaranteed-spread hunting without making directional claims.

---

Phase 4 - Entry Price Analysis

Three anomalous bands stand out:

1. $0.00-$0.10 at +23.5% ROI: The lowest-probability entries return the highest ROI. These are longshot hedge-side buys (the "No" side of an event where the dominant side is already heavily priced as favorite). When the longshot wins, the payout is enormous. The 6.2% win rate at sub-$0.10 is approximately calibrated to the implied probability, but the ROI is outsized because wins pay 10-100x.

1. $0.30-$0.40 at +18.5% ROI: This band captures many of the "spread lock" buys where the bot enters at near-even prices and both sides land in the $0.30-$0.70 zone. The high ROI here likely reflects clean spread execution where paired costs were well below $1.00.

1. $0.60-$0.70 at -3.2% ROI: The mild-favorite zone is the worst band by ROI. Entries here often represent the "late hedge" scenario - buying the second side after the market has moved toward the dominant outcome, creating paired costs above $1.00 on those specific fills.

Sub-bucket concentration check: The price histogram is broad across all 101 price points with no single penny dominating. The $0.50 band is the single most-traded cent, and common prices like $0.47, $0.51, $0.57, $0.59, $0.63, $0.69, $0.71, $0.73, $0.77 appear frequently in the CSV (reflecting rounded lines on baseball over/unders and spreads). This is a multi-tick, multi-market strategy, not a single-price anchor.

---

Phase 5 - Category and Vertical Breakdown

Tennis at 129,770 trades is the single most-traded vertical - nearly triple the next category (MLB at 52,593). The 33.9% tennis win rate reflects the nature of paired-side tennis markets: buying both "Player A wins" and "Player B wins" on 3-set matches where one player is a strong favorite means the underdog ticket wins roughly 30-40% of the time. The spread on tennis is typically tighter than on major US sports, but the volume more than compensates.

MLB over/under and spread markets are heavily represented in the CSV tail. The bot runs multi-line coverage on games: for the same mlb-mil-chc-2026-05-19 event, it simultaneously has positions in the moneyline, O/U 6.5, O/U 7.5, O/U 8.5 markets. This multi-line approach multiplies spread capture opportunities per event.

Best single markets by P/L:

The Athletics vs. Angels market at +$121,624 on only $32K of volume is a +376% ROI on a single game. This is not spread capture - it is a large directional bet that paid off. The 134 trades and 55 wins (41% WR) with that payoff suggests the bot went heavily long on one outcome at low implied probability and hit.

---

Phase 6 - Timing and Execution Analysis

Hourly P/L distribution:

Unlike directional traders with clear peak/sleep windows, this bot runs 24/7 - the hourly trade counts range from 2,939 (hour 5) to 14,414 (hour 0), with no zero-trade hours. The P/L variance by hour is large but not driven by volume - hours 00:00 and 01:00 both have high trade counts but are the worst P/L hours. This likely reflects late-game NBA activity where the bot's basketball model underperformed.

Day-of-week breakdown:

Friday is the only negative day at -2.15% ROI. Sunday is the standout at +6.83% ROI and +$426,595 absolute. The week-1 loss (-$204,469) then recovery to +$1,136,314 cumulative is the dominant pattern in the rolling windows.

The second-side lag (836 seconds median, ~14 minutes) is the most important execution characteristic. The bot is not an instantaneous both-sides entry bot. It enters the first side, watches the market, then hedges when the second-side price looks attractive. This creates paired cost variance across positions within the same market.

---

Phase 7 - Filter Experiments

The high-conviction filter is the most important finding: restricting to markets where dominance ratio exceeds 2.0x and taking only the dominant leg turns +2.60% baseline ROI into +8.82% ROI on $21.5M deployed. This is a near-doubling of ROI by simply skipping the near-even spread plays and focusing on the directional signal.

Full filter commentary appears in the Filters tab.

---

Phase 8 - Rolling Window Consistency

The rolling window pattern is notably front-loaded with losses: the first seven rolling-7 windows (April 20-26) are all negative, with the rolling-7 P/L bottoming at -$282,171 on April 23. The strategy then recovers strongly through weeks 3-5, with the rolling-15 crossing positive on April 29 and remaining positive through the rest of the window.

Weekly P/L progression:
  Week 17 (Apr 20-26):  -$204,469   Cumul: -$204,469
  Week 18 (Apr 27-May 3): +$789,032  Cumul: +$584,563
  Week 19 (May 4-10):  +$65,872    Cumul: +$650,436
  Week 20 (May 11-17): +$439,588   Cumul: +$1,090,024
  Week 21 (May 18-19): +$39,044    Cumul: +$1,129,068

The Week 1 loss of -$204K coincides with the peak of NBA playoff activity (Magic vs. Pistons, early NBA series where the bot appeared to lose significantly on spread positions that moved against it). The recovery in Week 2 (+$789K) is the single largest weekly gain and likely reflects tennis and MLB volume catching up plus the NBA model stabilizing.

The 76.7% rolling-window green rate is lower than elite directional traders (SirMartingale was 100%), but is expected for a spread-capture strategy that can absorb multi-day losses when a category misfires, then recover when volume cycles back to profitable categories.

---

Phase 9 - P/L Decomposition

The spread P/L of +$1,776,733 is the guaranteed floor from all sub-$1.00 paired positions. But the hedge tax of -$16.5M is the aggregate cost of all losing-side positions - this is not a "cost" in the traditional sense but rather the mechanical consequence of having settled $0 on the loser of every matched pair. The strategy works because the spread between paired cost and $1.00 (the guaranteed win) exceeds the proportional outflow on losing sides.

The +$912,829 implied directional residual represents the incremental P/L from: (a) markets where the bot took a dominant-side position without hedging, (b) markets where the dominant tilt was large enough that winning the dominant side generated more than the hedge tax consumed, and (c) the exceptional individual directional hits like Athletics vs. Angels (+$121K) and Yankees vs. Brewers (+$71K).

There is no SELL-leg decomposition because the wallet has zero SELL trades. All exits are at settlement.

---

Phase 10 - Strategy Specification

One-sentence summary: A 24/7 automated both-sides spread-capture bot across sports prediction markets that simultaneously buys YES and NO on the same game at sub-$1.00 combined cost, with a directional model that skews capital allocation toward the predicted winner and generates incremental alpha at 3x+ tilt ratios.

Edge sources (stacked):

1. Structural spread capture - buying both sides at combined cost below $1.00 guarantees profit at settlement regardless of outcome. Median locked spread of $0.0252 per unit across 4,202 both-sides markets.
2. Directional alpha on tilted positions - dominant-side win rate of 59.6% at 3x+ tilt confirms a real predictive model. This adds ~9.6 percentage points of win-rate alpha above the coin-flip baseline.
3. Multi-line event coverage - running positions on moneyline + multiple O/U totals + spread markets for the same event multiplies spread opportunities per game without proportional model risk.

Primary risk factor: Second-side lag (836s median) creates adverse selection on hedges in fast-moving markets. The 40.2% of paired markets with cost above $1.00 are guaranteed losses on those sub-pairs, partially offsetting the guaranteed wins on the 59.8% below $1.00.

What replicators must do: See Playbook tab for the full implementable specification.

Wallet: 0xfe787d2da716d60e8acff57fb87eb13cd4d10319 Window: 2026-04-20 → 2026-05-19 (30 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: 55.09% (4,202 of 7,628 markets)
• Median paired cost: $0.9748
• Mean paired cost: $0.9748
• Paired cost % under $1.00: 59.8%
• Paired cost % under $0.97: 47.9%
• Median 2nd-side hedge lag: 836s

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: 21 of 30 (70.0%)
• Rolling 7-day P/L range: -$282,171 → +$793,225
• Rolling 15-day windows green: 22 of 30 (73.3%)
• Rolling 15-day P/L range: -$282,171 → +$905,151

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 2026-05-21 11:00 UTC.

Wallet: 0xfe787d2da716d60e8acff57fb87eb13cd4d10319 Window: 2026-04-20 to 2026-05-19 Baseline: 233,145 resolved BUYs · 43.2% WR · $43.6M deployed · +$1,136,314 P/L · +2.60% ROI

The headline result

One filter produces massive, genuine lift. The category filter is narrow but useful. The hour filter slightly hurts. The price filter helps moderately.

The high-conviction dominance filter is the dominant finding: restricting to only the 2.0x+ dominance bucket and taking the dominant side only turns +2.60% ROI into +8.82% ROI on $21.5M of capital. This is the single most important lever available to a replicator. The strategy's directional signal is concentrated at high dominance; the near-even spread plays dilute the overall ROI substantially.

The price band filter ($0.30-$0.70) provides moderate lift (+1.03pp to +3.63%). The combined filter - high conviction plus top category plus hour exclusion - reaches +25.8% ROI but on a very small sample ($44K, 120 trades) that is not statistically representative.

---

Filter results table

---

Filter-by-filter commentary

1. High-conviction filter (dominance ≥ 2.0x, dominant leg only) → MEANINGFUL_LIFT

This is the single most valuable filter in the battery. Restricting to markets where the bot's own capital allocation reveals dominance ≥ 2.0x and taking only the dominant leg returns +$1,896,174 on $21.5M deployed (+8.82% ROI) versus the baseline of +2.60%. The lift of +6.22 percentage points is real and large.

The mechanism: at 2.0-3.0x dominance, the dominant side wins 58.8% of the time; at 3.0x+, it wins 59.6%. Both substantially exceed the 43.2% overall win rate. The bot's own sizing is functioning as a signal - the more capital it tilts toward one side, the more likely that side is to win.

The filter also raises the P/L in absolute terms: +$1,896,174 vs. +$1,136,314 unfiltered. This seems paradoxical (filtering out positions increases absolute P/L) but makes sense: the low-dominance positions include many near-even spreads where the hedge tax eats gains, and single-sided positions in losing categories (especially NBA) that drag down the total. The high-conviction filter implicitly excludes most NBA blowup exposure.

For a replicator: this filter is the core of the investable strategy. The near-even spread plays (1.0-1.5x dominance) generate guaranteed-spread P/L but at thin ROI; the high-dominance directional plays are the real alpha source. A replicator with limited capital should run only the 2.0x+ bucket.

2. Price band filter ($0.30-$0.70) → MODERATE_LIFT

The standard sweet-spot filter (entries between $0.30 and $0.70) improves ROI from +2.60% to +3.63% on $27.2M deployed (+$987K P/L). The lift is +1.03 percentage points on a substantial capital base.

The mechanism: the $0.60-$0.70 band is slightly negative (-3.2% ROI) and the $0.70-$0.80 band is slightly negative (-0.9%). The filter excludes these bands. However, the sub-$0.10 band at +23.5% ROI is also excluded, which costs some longshot upside. The net effect is modest positive lift.

This filter is more useful as a risk management tool than a pure ROI maximizer: it caps exposure to the potentially mispriced late-hedge positions (which tend to land in the $0.60-$0.80 band when the market has moved toward the winning side).

Caveat: the $0.30-$0.40 band at +18.5% ROI is included by this filter and is clearly a good band. The $0.10-$0.20 band at -4.8% ROI is correctly excluded. The filter is correctly designed for this strategy's price structure.

3. Top category filter (UFC/MMA only) → NOT_APPLICABLE_AT_SCALE

UFC/MMA returns +35.0% ROI on $86,456 deployed (230 trades). This is extraordinary, but the sample is too small to be actionable on its own. The 230 trades across 30 days (7.7 fights per day average) represent the full available UFC/MMA market volume on Polymarket during the window - you cannot increase exposure to this category beyond what the market offers.

As a portfolio instruction rather than a filter: weight UFC/MMA at maximum available capital whenever it is on the menu. The +35% ROI likely reflects the bot's model being particularly well-calibrated for fight sports, but the absolute P/L contribution (+$30K) is a rounding error on the total book.

The combined filter that stacks high-conviction + UFC/MMA + hour exclusion reaches +25.8% ROI but on only $44K and 120 trades. This is a statistical artifact of stacking narrow filters on a small overlap rather than a genuine investable signal.

4. Hour exclusion filter (drop hours 6, 9, 11, 14) → DESTRUCTIVE

Excluding the four worst-performing hours (6, 9, 11, 14 UTC) actually reduces total P/L from +$1,136,314 to +$817,634 and ROI from +2.60% to +2.12%. This is a -0.48pp drag, not a lift.

The mechanism: the "worst hours" by win rate are not the worst hours by absolute P/L. Hours 6 UTC (3.5% WR) and 9 UTC (30.4% WR) have low win rates but are not high-volume hours (3,500 and 11,915 trades respectively). The filter removes substantial good-ROI capital from the excluded hours alongside the bad fills.

More importantly, hours 6, 9, 11 overlap with morning tennis and early baseball activity - removing them cuts into the Tennis and MLB spread positions that are the engine of the profitable categories.

The four identified "worst hours" by win rate (6, 9, 11, 14 UTC) have win rates of 34.9%, 30.4%, 33.1%, and 35.7% respectively. These low win rates are partially explained by the spread structure: hours with high tennis volume will naturally have lower win rates because tennis spread buys on heavy favorites (sub-$0.30 tickets) drag the overall WR down even as they contribute positive spread P/L.

Do not apply the hour exclusion filter to this strategy. The win-rate-based hour ranking is misleading for a spread-capture book.

5. Category exclusion (drop NBA) → Hypothetical Analysis

The filter battery did not explicitly test NBA exclusion, but the data is clear. NBA generated -$191,047 on $7.9M deployed (-2.47% ROI). Excluding NBA from the baseline would improve total P/L to approximately +$1,327,361 on $35.7M deployed (+3.72% ROI).

The NBA underperformance was systematic and concentrated in specific matchups (Timberwolves-Nuggets, Cavaliers-Raptors, Rockets-Lakers second instance, Hawks-Knicks). This suggests the bot's basketball model was poorly calibrated for these specific series - possibly a data input problem or a model parameter that worked in regular season but not in playoff format.

For a replicator: consider excluding NBA during the playoffs or running NBA at reduced dominance thresholds (only deploy at 3x+ dominance rather than 2x+ for basketball). The regular-season baseball and tennis spreads are where the reliable edge lives.

---

The actionable filter stack

If forced to name a single filter combination that a replicator should apply:

The hour filter is the clearest "don't apply" instruction. The high-conviction filter is the clearest "always apply" instruction. The NBA exclusion during playoffs is judgment-based but strongly supported by the data.

The combined filter with these three stacked (high-conviction + no NBA playoff + price $0.30-$0.70) would likely produce ROI in the +10-12% range on roughly $15-20M of capital - roughly a 4x improvement over the raw baseline strategy, using the bot's own sizing decisions as the primary selection signal.

Source wallet: 0xfe787d2da716d60e8acff57fb87eb13cd4d10319 Strategy: Both-sides spread capture with directional dominance tilt across sports prediction markets Reference book: $43.6M BUY notional · +$1,136,314 net P/L · +2.60% ROI in 30 days (baseline); +8.82% ROI under high-conviction filter

---

One-paragraph operator brief

Build a Polymarket bot that monitors live sports prediction markets across tennis, MLB, soccer, and UFC/MMA - and buys both YES and NO on the same game whenever the combined bid-ask spread locks in a paired cost below $1.00. Simultaneously run a game-outcome model to determine which side has the better expected probability. If your model assigns one side more than 2.0x the capital weight, that is your dominant side: size it proportionally larger and still hedge the other side. Enter the dominant side first. Hedge within 15 minutes or skip the hedge if the second-side price has moved past $1.00 combined. Apply the strategy only to markets where your model has confidence; skip NBA during post-season unless dominance exceeds 3.0x. Expect +2-4% ROI on deployed capital at baseline; expect +8-10% ROI if you filter to dominance ≥ 2.0x dominant-leg-only fills.

---

1. Market selection

Why tennis is the core: Tennis has the highest trade count (55.6% of the book), consistent opponent matchup quality, and relatively efficient spreads. The 3.4% ROI on $17.2M across 30 days confirms stable, repeatable edge. Tennis also has no playoff/format complications that corrupt the model.

Why MLB is the secondary: Multiple market types per game (moneyline + 5+ O/U lines + spread variants) multiply the spread opportunities without increasing event exposure. The CSV shows the bot running 8+ simultaneous positions on mlb-mil-chc-2026-05-19 - all drawing from the same game model.

Why NBA deserves caution: The playoff format creates asymmetric information flow (coaching adjustments, injury news, line-up changes) that a market-wide model struggles to incorporate. The -$191K NBA loss is the single largest drag on the book. Limit NBA to regular season or require 3.0x+ dominance before deploying.

---

2. Entry logic

Phase 1: Dominant-side entry

def should_enter_dominant(market, model):
    # Category whitelist
    if market.category not in ("Tennis", "MLB", "Soccer", "UFC/MMA"):
        if market.category == "NBA" and model.dominance(market) < 3.0:
            return None
    
    # Liquidity check
    if market.best_ask_depth_usdc < min_clip_size:
        return None
    
    # Model must have a view
    dominant_side, confidence = model.predict(market)
    if confidence < 0.50:
        return None  # no signal, skip
    
    return dominant_side

Phase 2: Hedge entry (within 15-minute window)

def should_hedge(market, dominant_fill, model):
    hedge_side = opposite(dominant_fill.outcome)
    hedge_ask = market.best_ask(hedge_side)
    
    # Compute implied paired cost
    dominant_vwap = dominant_fill.avg_price
    paired_cost = dominant_vwap + hedge_ask
    
    if paired_cost >= 1.00:
        return None  # spread locked above $1.00 - skip hedge or reduce
    
    # Hedge window: must enter within 15 minutes of dominant fill
    if (now() - dominant_fill.timestamp) > 900:
        return None  # window expired, let dominant leg ride as pure directional
    
    return hedge_side, hedge_ask

---

3. Sizing model

Two-tier sizing based on dominance ratio:

dominance_ratio = usdc_planned_dominant / usdc_planned_hedge

If dominance_ratio < 1.5:   # Near-even spread play
    dominant_clip = base_clip
    hedge_clip    = base_clip

If 1.5 <= dominance_ratio < 2.0:  # Mild directional lean
    dominant_clip = base_clip * 1.5
    hedge_clip    = base_clip

If 2.0 <= dominance_ratio < 3.0:  # Meaningful signal
    dominant_clip = base_clip * 2.5
    hedge_clip    = base_clip

If dominance_ratio >= 3.0:  # High conviction
    dominant_clip = base_clip * 5.0+
    hedge_clip    = base_clip

Reference scale (matching observed book):

Important: The reference book shows a $190,399 single fill - an extreme outlier that is 38× the P99 of $3,335. Do not treat this as a normal operating parameter. This was likely a specific event (the Athletics vs. Angels market at +$121K P/L) where the operator made a large directional bet outside the standard spread playbook. Exclude from your sizing model calibration.

The power-law size distribution (top 5% of trades hold 71.6% of capital) is real and structural: most fills are small probes that complete the multi-line coverage, while the occasional large clip is where the model has exceptional conviction. Design your sizing model to allow for 10-50× base-clip positions on 3x+ dominance markets.

---

4. Paired-cost discipline

This is the core rule that separates profitable spread capture from undisciplined betting:

For each market where you hold both sides:

paired_cost = YES_VWAP + NO_VWAP
              (weighted average across all fills on each side)

If paired_cost < 0.97:  guaranteed_profit = 1.00 - paired_cost >= 3¢/unit
If 0.97 <= paired_cost < 1.00: thin spread - reduce size next time
If paired_cost >= 1.00: guaranteed loss - post-mortem this market

Target metrics (matching reference book):

• Median paired cost: ≤ $0.975
• % markets below $0.97: ≥ 45%
• % markets below $1.00: ≥ 60%

The 40.2% of both-sides markets with paired cost above $1.00 in the reference book represent failures of the hedge timing. Most of these occur when the second-side price moved significantly before the hedge executed. The 836-second median lag is too long for fast-moving markets. A replicator should target a tighter hedge window (under 300 seconds) to reduce this fraction.

---

5. Dominance-ratio calibration

The high-conviction filter finding (dom ≥ 2.0x produces +8.82% ROI vs. +2.60% unfiltered) is the most important calibration data point. A replicator running a capital-constrained book should:

1. Always deploy at 2.0x+ dominance markets (max capital within liquidity)
2. Deprioritize 1.0-1.5x markets in capital allocation
3. Never skip a 3.0x+ market - even if it requires borrowing from the 1.0-1.5x allocation

---

6. Category priority and scheduling

Day-of-week adjustment: Sunday is the highest-ROI day (+6.83%). Saturday is second (+3.80%). Monday is third (+3.49%). Friday is the only negative day (-2.15%). Consider running at full capacity Sunday through Thursday and monitoring closely on Fridays (NBA Friday games appeared to drive Friday losses in the reference period).

Hourly P/L note: Hours 18:00 (+$232K), 23:00 (+$213K), 14:00 (+$188K), and 10:00 (+$183K) are the four best UTC hours by absolute P/L. Hours 01:00 (-$183K) and 03:00 (-$147K) are the worst. However, do not filter by hour - the hour-exclusion filter was destructive in testing. Use this data for monitoring rather than filtering.

---

7. Multi-line event coverage

One of the most capital-efficient features of this strategy is multi-line coverage: deploying on multiple betting lines for the same underlying game.

For an MLB game, the reference book runs simultaneously on:

• Moneyline (Team A wins)
• O/U 5.5 total runs
• O/U 6.5 total runs
• O/U 7.5 total runs
• O/U 8.5 total runs
• O/U 9.5 total runs
• Spread variants (Team A -1.5, Team B -1.5)

Each of these is a separate Polymarket market with its own orderbook and liquidity. The bot's game model generates probability estimates for each market independently. The paired-cost rule applies to each market separately - a sub-$1.00 paired cost on the O/U 7.5 market is independent of the paired cost on the moneyline.

Implementation rule: For every game event, enumerate all active Polymarket markets and filter to those where your model has a view AND the paired cost can be locked below $1.00. Deploy the standard sizing on each qualifying market. The incremental cost per additional line is low (the model already has the game prediction), and the incremental spread P/L is additive.

---

8. Risk management

Portfolio-level drawdown rule: If any rolling 7-day P/L goes negative by more than -1% of deployed capital, pause new position initiation in the loss category and review model inputs. The reference book's Week 1 loss of -$204K (-0.47% of the month's total deployed) recovered fully by Week 2; a -1% weekly threshold would not have triggered a pause in that case, which is appropriate.

Per-market cap: No single event should receive more than 1% of monthly deployed capital without manual review. The reference book violated this on a few marquee events (Magic vs. Pistons appeared to have $15K+ deployed), and the outsized losses in those games contributed to the Week 1 drawdown. Mechanical position sizing prevents this.

---

9. Operational requirements

---

10. Diagnostic checklist: is the bot still working?

Run weekly:

---

What this playbook deliberately does NOT include

• No hour-of-day exclusion filters. The hour-exclusion test was destructive in this dataset. Run 24/7 and let the category model carry the timing selection.
• No single-category focus. UFC/MMA at +35% ROI sounds compelling but the absolute capacity is too small ($86K/month) to matter for a real book. Tennis + MLB is the actual alpha engine.
• No fixed $0.30-$0.70 price band restriction. The $0.00-$0.10 band at +23.5% ROI is load-bearing (representing genuine longshot hedge wins). Filtering it out reduces P/L.
• No pure directional play (no hedge). The strategy's structural edge is the guaranteed spread, not the directional call alone. Running as a pure directional bettor without the hedge reduces Sharpe even if it occasionally improves raw returns.
• No hold-to-resolution-only model. This strategy has zero sells in the reference period, but this is by design: you are buying both sides and one side will always win at $1.00. Adding an early-exit SELL mechanism is possible but changes the spread-capture mechanics entirely - it becomes a different strategy.
• No NHL expansion. The reference book ran -2.23% ROI on NHL with 3,203 trades. The market structure on hockey totals and spreads does not appear to generate reliable spread opportunities at this operator's model quality. Avoid NHL until a specific edge is identified.

The core discipline is: enter both sides when the spread locks below $1.00, size proportional to your model's directional conviction, and do nothing else clever. Every "improvement" the operator apparently tested and rejected by the data (NHL expansion, hourly timing, single-category focus) shows up as drag when applied. Trust the data.