Ratue

Wallet: 0x912a58103662ebe2e30328a305bc33131eca0f92 Window: 2026-04-08 to 2026-05-05 (28 calendar days, 28 active) Universe: 284,661 trades across 7,674 markets, $987,333 gross BUY notional, 0 SELL trades Net P/L: +$18,952 on $987,333 deployed = +1.92% ROI (resolved-BUY methodology)

The Punchline

Ratue is a high-volume both-sides spread-capture bot operating exclusively on short-duration crypto Up/Down markets. Every fill, every day, for 28 consecutive days, the bot enters both the Up and Down sides of 15-minute BTC, ETH, and SOL windows and collects the spread at resolution. There are zero SELL trades. Every position rides to settlement. The P/L comes from buying both sides for less than $1.00 in aggregate and collecting exactly $1.00 when the market resolves.

The strategy is not glamorous but it is mechanically sound. The mean paired cost across 6,775 dual-sided markets is $0.9450, meaning the bot spends on average 94.5 cents to guarantee a $1.00 payout. The 5.5-cent average spread, multiplied across $987K of capital and 284,661 fills, produces the gross spread P/L of $20,611. After deducting $196,870 in hedge-tax losses (the mandatory losing-side capital in each paired market), the net realized P/L is +$18,952. The math is unglamorous but the operation is consistent: 4 of 5 rolling weekly windows are profitable, 27 of 28 rolling 7-day windows are positive, and the cumulative P/L climbs every week for the first four weeks of the observation period.

Layered on top of the spread-capture base is a real directional signal. In markets where Ratue puts 3x or more capital on one side, that side wins 85.2% of the time. The sizing ratio is not random - it is a genuine edge expression that amplifies returns beyond what pure spread capture alone would produce.

What He Trades

The universe is three assets, one duration:

BTC Up/Down 15m    bulk of volume    slug: btc-updown-15m-*
ETH Up/Down 15m    significant       slug: eth-updown-15m-*
SOL Up/Down 15m    meaningful        slug: sol-updown-15m-*

Every single one of the 7,674 markets is a 15-minute crypto price window. No sports, no politics, no hourly or 5-minute windows, no 4-hour windows. The portfolio is a single-category, single-duration book. The CSV confirms the three-asset structure: every market slug in the sample is btc-updown-15m-*, eth-updown-15m-*, or sol-updown-15m-*.

The top markets by volume confirm that BTC 15m windows dominate, but ETH and SOL appear regularly. The best single market by P/L is Ethereum Up or Down - April 29, 4:00PM-4:15PM ET with +$593.77 on $76.67 deployed - a +774% return on a very small capital commitment, suggesting the bot found an extreme mispricing on that market. The worst single market is Bitcoin Up or Down - April 12, 7:30PM-7:45PM ET with -$683.81 on $683.81 deployed and 0 wins, meaning the bot bought the losing side exclusively (or the paired cost was well above $1.00).

The 88.3% both-sides rate (6,775 of 7,674 markets) is the single most important structural fact about this wallet. Markets with only one side purchased (899 markets, 11.7%) represent probe fills, partial executions, or rare directional-only bets.

The Order of Operations - One Market, Trade by Trade

The cleanest illustration of the strategy is Bitcoin Up or Down - April 7, 8:00PM-8:15PM ET (the first market at the start of the CSV tail). The bot enters this market across the full 15-minute window, accumulating on both sides simultaneously.

Walk-through:

The market opens at approximately 8:00PM ET. The bot immediately begins accumulating both sides. In the early minutes it probes the Down side at $0.27-$0.31 and the Up side at $0.39-$0.51. These are coin-flip zone bids - the Up side trades at 40-50 cents, implying roughly even odds.

As the 15-minute window progresses and price information arrives (from the underlying BTC spot price), the bot shifts dramatically. In the final 2-3 minutes (00:12 to 00:14 UTC), the bot identifies that BTC has moved and the Up side is now pricing at $0.88-$0.89. It dumps large clips into the Up side: 12.30 shares, 13.30 shares, 5.06 shares, 7.91 shares - all at $0.88-$0.89. Simultaneously it still takes Down side fills at $0.11-$0.17 (the complementary cheap side).

The paired cost in the final phase: Up at $0.89 + Down at $0.11 = $1.00. That is not spread capture, that is fair value. But the early-window Down fills at $0.27-$0.31 are the alpha: the bot paid 27-31 cents for shares that were never going to pay off (BTC went Up), but it also paid 13-16 cents for Up shares in the early probe phase when the true probability was already skewing Up. The early Up fills at $0.13-$0.16 that survive to resolution at $1.00 each generate a 6-7x return.

The strategy is: buy both sides early at wide spreads, accumulate the side that is moving into favorable probability territory, and use the cheap side as a paired-cost anchor. The bot never knows which way BTC will go at the start of the window; it finds out as the window progresses and adjusts its allocation accordingly.

Why It Works - The Math

The spread-capture logic:

For each paired market:
  paired_cost = (USDC spent on Up side / Up shares acquired)
              + (USDC spent on Down side / Down shares acquired)
  spread_profit = (paired shares at resolution) x (1.00 - paired_cost)

Aggregate across 6,775 paired markets:
  Mean paired cost:          $0.9450
  Median paired cost:        $0.9999  (but mean is pulled down by low-cost pairs)
  % with paired cost < $1.00: 50.0%
  % with paired cost < $0.97: 45.2%
  Gross spread P/L:          +$20,611
  Hedge tax:                 -$196,870
  Net realized P/L:          +$18,952

The hedge tax is the dominant cost. In each paired market, the losing side's capital is fully forfeit. The bot spends ~$197K on shares that resolve to $0. The spread P/L of $20K only covers the net cost of running the strategy because:

1. Paired costs below $1.00 generate guaranteed profit proportional to 1 - paired_cost.
2. The directional skew on high-dominance markets means more capital sits on the winning side, reducing the effective hedge tax per dollar deployed.

The directional layer EV:

Dominance bucket    Count    Dom WR    Implied edge
1.0-1.5x            943       55.5%    +5.5% over 50%
1.5-2.0x            656       62.9%    +12.9% over 50%
2.0-3.0x            927       71.8%    +21.8% over 50%
3.0x+              4,249      85.2%    +35.2% over 50%

The 3.0x+ bucket contains 4,249 markets - 55.3% of all paired markets - and shows an 85.2% dominant-side win rate. Whatever signal drives the sizing allocation is accurate at a level that cannot be attributed to the spread structure alone. The spread structure guarantees a floor return; the directional signal provides the majority of the upside.

The total EV decomposition confirms this. If the bot had sized every market equally (1:1 on both sides), the spread P/L of $20,611 would still exist. The net P/L would be lower, not higher, because the directional skew redirects capital to the winning side and reduces hedge tax per resolved winner. The directional layer adds alpha on top of the structural spread.

Phase 1 - Trader Profile

Scale and activity: - 284,661 BUY trades, 0 SELL trades in 28 days (all 28 days active) - $987,333 BUY notional, $0 SELL notional - 7,674 unique markets, 7,674 unique events (one event per market) - Average 10,167 fills per day, 7.27 seconds per fill at the median

Trade size distribution:

The distribution is highly right-skewed. The median fill is $1.43, but the mean is $3.47, and the top 5% of trades carry 48.7% of the capital. The max is $150, which is 105x the median. This is a power-law size distribution consistent with a strategy that scales position size by conviction (dominance ratio) rather than by fixed clip. When the bot is highly confident in one side, it fires large clips; when probing, it fires tiny clips.

Execution signature: - Median inter-fill gap: 2.0 seconds - 70.0% of consecutive fills under 10 seconds - 90.4% of consecutive fills under 60 seconds - 100% of fills under 1 hour (no orphaned positions)

The 2-second median gap and 70% sub-10-second rate confirm a fully automated bot. Human traders do not sustain 2-second median fill cadences across 284,661 fills over 28 days. The gap statistics are slightly longer than a pure HFT bot because the fills are spread across three assets and many concurrent markets - the bot is managing dozens of open positions simultaneously and filling them in rotation.

Active hours: Uniform. Trades in every single hour of the day. No sleep window.

The distribution is nearly flat across all 24 hours. The busiest hours are 00:00 (13,213 fills) and 12:00-13:00 (~12,700 fills each), but no hour has fewer than 10,643 fills. This is a 24/7 automated system with no operator sleep constraint. This is the starkest operational difference from SirMartingale, which has a clear 3-hour overnight dead zone.

Archetype: SPREAD CAPTURE + DIRECTIONAL SKEW 88.3% both-sides participation, conviction-scaled sizing, no SELL leg.

Phase 2 - Core Strategy Identification

Both-sides participation rate: 88.3% (6,775 of 7,674 markets). This is definitionally a market-making / spread-capture strategy. A both-sides rate above 60% per ANALYSIS_SPEC Phase 2 confirms spread capture as the primary archetype.

The paired cost profile confirms it: - Mean paired cost: $0.9450 (5.5 cents expected profit per $1 guaranteed) - Median paired cost: $0.9999 (half of paired markets have cost within 0.01 cents of $1.00) - 50.0% of paired markets have cost under $1.00 - 45.2% have cost under $0.97

The median is close to $1.00 because many pairs are made at near-fair-value prices (late-window fills when the market has already priced the outcome accurately). The mean pulls down to $0.9450 because the bot finds meaningful sub-$1.00 pairs consistently. The wide spread on the paired cost distribution reflects the full range of market conditions: some 15-minute windows are deeply mispriced (paired cost $0.20-$0.50, early in the window), others approach fair value (paired cost $0.95-$1.00, late in the window).

Classification:

Primary: A (Both-Sides Spread Capture). 88.3% both-sides rate, guaranteed spread locked in when paired cost < $1.00.

Secondary: B (Directional Betting with conviction signal). The 3.0x+ dominance bucket with 85.2% dominant-side win rate is real directional edge, not noise. The sizing allocation encodes genuine information about the direction of price movement within each window.

The combination is what makes this strategy distinct from a pure market maker. A pure market maker buys both sides in equal amounts and collects only the spread. Ratue also buys both sides, but it sizes them asymmetrically based on a directional signal - and that signal is accurate enough to generate meaningful additional P/L beyond the spread.

Not: - Directional-only: the 88.3% both-sides rate rules this out - Latency arbitrage: no SELL trades, 24/7 operation, broad price band entry - Copy trading: fills on market open, not following another wallet - DCA accumulator: multi-second fills within single 15-minute windows, not slow accumulation

Phase 3 - Dominance Ratio Analysis

The monotonic increase in dominant-side win rate from 55.5% at low dominance to 85.2% at high dominance is the key finding. A dominant-side win rate of 85.2% across 4,249 markets (the largest single bucket) is statistically overwhelming. This cannot be explained by the paired-cost structure alone.

The average paired cost is $0.978 in the 3.0x+ bucket - nearly fair value. This means the bot is not extracting much spread profit in high-conviction markets; it is expressing a directional bet under the cover of a paired structure. The spread in those markets is nearly zero; the alpha is almost entirely directional.

Conversely, the 1.0-1.5x bucket has an average paired cost of $0.865 - a 13.5-cent spread - but only 55.5% dominant-side accuracy. Those markets are where the spread P/L is highest per market, but the directional signal is weakest.

Interpretation: The bot has two modes that it runs simultaneously. In markets where it detects strong directional signal, it loads up on one side (3.0x+ dominance) and accepts near-zero spread for the directional return. In markets where it is directionally uncertain, it runs a balanced paired structure with lower dominance and higher spread.

Phase 4 - Entry Price Analysis

The price distribution is bimodal. The largest capital concentration is $0.90-$1.00 at $557K (56.4% of all capital), representing the expensive/winning side of each paired market. The second-largest by trade count is $0.00-$0.30 (130,868 fills), representing the cheap/losing side of each pair.

The sub-bucket concentration check is critical. The $0.90-$1.00 band dominates absolutely - $557K of $987K total. Within this band, the bot frequently buys at $0.91-$0.99 (the winning side of high-conviction pairs). The sub-$0.10 band (46,245 fills, $24,501 spent) generates +22.8% ROI because these are cheap-side buys that occasionally win when the bot is wrong about direction - paying $0.04-$0.09 for something that resolves at $1.00 produces enormous per-share returns on rare wins.

The $0.30-$0.40 band is the only band with negative ROI (-0.7%, -$434). This is the coin-flip zone where the bot pays mid-range prices for shares that win at roughly the implied probability - no spread, no directional edge, no excess return. This confirms the strategy avoids pure mid-band directional betting and clusters at the extremes (very cheap paired legs or very expensive winning sides).

Phase 5 - Category and Vertical Breakdown

Only one category exists: Crypto, 100% of all 284,661 trades and $987,333 deployed.

The 39.2% win rate looks bad in isolation but is correct for a paired strategy. In any market where the bot buys both sides and the Up side wins, the Down side trades are losses and vice versa. The overall win rate of 39.2% reflects the structural reality that approximately half of all fills are on the losing side - but the strategy is designed to profit from the spread between both sides, not from picking winners.

By asset (from slug analysis across the CSV sample): - BTC 15m: dominant by volume, most markets - ETH 15m: significant secondary asset - SOL 15m: present but smaller (SOL appears in CSV tail sample at the early April 8 entries)

The best-performing individual markets skew toward ETH (e.g., "Ethereum Up or Down - April 29, 4:00PM-4:15PM ET" at +$593.77 on $76.67 deployed, an 874% return), suggesting ETH 15m windows have wider mispricings relative to capital deployed. The top 10 best markets by P/L include 4 ETH windows and 6 BTC windows; the top 10 worst are dominated by BTC.

Phase 6 - Timing and Execution Analysis

Hourly P/L:

The hourly P/L is volatile. The best single hour is 03:00 UTC (+$3,486) and the worst is 05:00 UTC (-$2,714). This variance suggests the bot does not have a specific time-of-day edge - it operates 24/7 and the P/L swings reflect which hours happened to contain good vs. bad BTC/ETH/SOL price action for its open positions.

Day-of-week P/L:

Sunday is the best day (+$5,651, +4.0% ROI). Consistent with the SirMartingale finding, weekend days (particularly Sunday) may have thinner competition from professional market makers, widening the spreads available for capture. Saturday is the second-highest volume day but lowest ROI, suggesting the bot runs harder on weekends but faces more competition.

Accumulation window: Each 15-minute market sees fills spread across the full window. The CSV shows the bot entering both SOL and BTC markets within the same second (e.g., 2026-05-05T23:59:34Z has both SOL and BTC fills simultaneously). The second-side lag median of 256 seconds (just over 4 minutes) confirms intentional pairing - the bot enters the second side of each market within 4 minutes of the first, well within the 15-minute window.

Burst patterns: Multiple fills on the same market in the same second. The CSV shows 4-6 fills on eth-updown-15m-1778024700 at 23:59:26Z simultaneously. This is a fan-out pattern consistent with walking the order book across multiple price levels simultaneously.

Phase 7 - Filter Experiments

The filter analysis for this trader has one dramatically positive result and one meaningless result.

The high-conviction filter (dom ≥ 2x) is the most important finding in this entire report. Applying a dominance filter of 2x or higher collapses the sample to 114,201 fills on $723,554 deployed and generates +$90,538 P/L - a +12.51% ROI versus 1.92% on the unfiltered book. The filter captures all the directional skew alpha while eliminating the low-conviction paired trades that contribute little P/L for their capital cost.

This is an extraordinary result. The high-conviction filter generates 4.78x more P/L per dollar deployed than the unfiltered strategy, on 73% of the capital. The delta is +$71,586 in absolute P/L terms.

The price filter ($0.30-$0.70) is destructive for the same reasons it would be for any spread-capture bot: it removes the expensive-side buys ($0.90+ zone) that are the winning legs of each pair, stripping the strategy of its settlement payouts.

See the Filters tab for per-filter commentary.

Phase 8 - Rolling Window Consistency

Every rolling 15-day window in the observation period is positive. The rolling 7-day has one near-zero window (April 21, +$131) : the single weakest 7-day is the April 21 window at +$131, barely above zero.

The weekly trajectory shows consistent improvement through the observation period:

Weeks 17 and 18 are significantly more profitable than weeks 15 and 16, suggesting the bot's directional signal sharpened over time (win rate improved from 38.7% in W15 to 42.2% in W18), or market conditions in late April were more favorable for spread capture.

Phase 9 - P/L Decomposition

The gross spread P/L of $20,611 exceeds the net P/L of $18,952, which means the hedge tax (the mandatory cost of the paired structure) is more than covered by the spread returns. The extra $1,659 in net P/L versus gross spread P/L comes from the directional skew - markets where the dominant side wins by more than the paired cost structure would predict.

This decomposition reveals the two-engine structure:

1. Spread engine: buys both sides for less than $1.00, collects guaranteed spread at resolution. Generates $20,611 gross.
2. Directional engine: asymmetric sizing toward the more-likely outcome, reducing effective hedge tax and adding directional P/L when the signal is correct.

The directional engine is the larger P/L driver in high-dominance markets. In 3.0x+ markets (85.2% dom WR, $0.978 avg paired cost), almost all the return is directional - the 2.2-cent spread barely matters, but the 35.2% excess win rate over random generates substantial additional P/L.

Phase 10 - Strategy Specification

One-sentence summary: A 24/7 automated spread-capture bot that buys both sides of every 15-minute BTC/ETH/SOL Up/Down window on Polymarket, locks in guaranteed spread when paired cost is below $1.00, and applies asymmetric sizing toward whichever side its directional signal favors.

Edge sources: 1. Structural spread capture: buying both sides for less than $1.00 guarantees profit at resolution 2. Directional accuracy: the sizing signal (3.0x+ dominance = 85.2% win rate) generates excess return beyond the spread

Universe: BTC/ETH/SOL 15-minute Up/Down markets exclusively. Zero category diversification.

Entry: Both sides of every active window. Entry timing spread across the full 15-minute window, not concentrated at open or close. Entry price ranges from $0.01 to $0.99, using full book depth.

Sizing: Conviction-scaled. Low-conviction markets: near-equal amounts on both sides (1.0-1.5x dominance). High-conviction markets: 3x+ more on the favored side, with the small counter-position as a paired anchor.

Exit: None - all positions held to settlement. No SELL trades in 28 days.

Risk: Max single-market loss bounded by total capital committed to that market. Worst single market in the window: -$684 (Bitcoin Up or Down - April 12, 7:30PM ET, 200 fills, 0 wins). Worst 5 markets: -$684, -$653, -$592, -$546, -$490. Loss events are bounded and infrequent.

Replication parameters: See Playbook tab.

Wallet: 0x912a58103662ebe2e30328a305bc33131eca0f92 Window: 2026-04-08 → 2026-05-05 (28 active / 28 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: 88.29% (6,775 of 7,674 markets)
• Median paired cost: $0.9999
• Mean paired cost: $0.9450
• Paired cost % under $1.00: 50.0%
• Paired cost % under $0.97: 45.2%
• Median 2nd-side hedge lag: 256s

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: 28 of 28 (100.0%)
• Rolling 7-day P/L range: +$131 → +$9,113
• Rolling 15-day windows green: 28 of 28 (100.0%)
• Rolling 15-day P/L range: +$432 → +$14,297

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 03:09 UTC.

Wallet: 0x912a58103662ebe2e30328a305bc33131eca0f92 Window: 2026-04-08 to 2026-05-05 Baseline: 284,661 BUYs, 39.2% WR, $987,333 deployed, +$18,952 P/L, +1.92% ROI

The headline result

One filter generates a transformational improvement. The rest are noise or destructive.

The high-conviction dominance filter (dom ≥ 2x, dominant side only) takes the strategy from +1.92% ROI to +12.51% ROI on 73% of the original capital. That is a 6.5x ROI improvement. In absolute P/L terms, the filtered strategy generates +$90,538 versus the unfiltered +$18,952 - a +$71,586 improvement. This is not a marginal refinement, it is a structural rearchitecting of where the money actually comes from.

Everything else in the standard filter battery is either meaningless (Crypto-only category filter), destructive (price band filter), already-optimized (no systematic hour-based improvement available given 24/7 operation), or counterproductive when combined with the dominance filter.

Filter results table

Filter-by-filter commentary

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

Applying the standard "sweet spot" price filter removes the $0.90-$1.00 zone that contains 56.4% of all capital ($557K) and most of the guaranteed winning-side settlement payouts. After filtering, only 101,567 trades and $184,824 of capital remain.

The result: P/L drops from +$18,952 to +$1,232, a -$17,720 damage. ROI also drops from +1.92% to +0.67%.

The mechanism is simple: this strategy operates at price extremes by design. The $0.90+ zone is where the bot buys the winning side of each pair. The sub-$0.20 zone is where it buys the cheap losing side. The mid-band ($0.30-$0.70) is the least profitable zone - these are the coin-flip region fills where neither the spread nor the directional signal is strongest. Filtering to include only the mid-band strips the strategy of its two primary alpha sources and leaves only the worst-performing segment.

Do not apply this filter to any spread-capture strategy. It is calibrated for directional bettors, not market makers.

2. High-conviction dominance filter (dom ≥ 2x, dominant leg only) MEANINGFUL LIFT

This is the most important filter finding in the entire analysis. Restricting fills to only the dominant side of markets where Ratue put at least 2x more capital on one side than the other generates:

• 114,201 fills (40.1% of total)
• $723,554 capital (73.3% of total)
• 70.6% win rate (vs 39.2% baseline)
• +$90,538 P/L (vs +$18,952 baseline)
• +12.51% ROI (vs +1.92% baseline)

The filter works because it isolates the fills where Ratue's directional signal is firing. From the dominance ratio table: - 2.0-3.0x dominance: 927 markets, 71.8% dom WR - 3.0x+ dominance: 4,249 markets, 85.2% dom WR - Combined: 5,176 markets, implied ~81% dom WR

The capital in these dominant-side fills ($723K of $988K total) is the capital carrying real directional conviction. The non-dominant side fills in these same markets (the smaller paired legs) are the hedge tax - they lose, but they lose small amounts relative to the large dominant-side wins.

This filter is immediately actionable. A replicator who copies only the dominant-side buys from markets where the dominance ratio exceeds 2x would extract 4.78x more P/L per dollar deployed than copying the full strategy.

3. Category filter (Crypto only) NO-OP

The entire book is Crypto. There are zero trades in any other category. Applying the category filter returns exactly the baseline result. The filter is structurally inapplicable to a single-category book.

This is not a weakness of the strategy - it is by design. The spread-capture logic requires 15-minute windows with complementary Up/Down structure, which currently exists only in the crypto price prediction market vertical.

4. Hour exclusion filter (exclude hours 5, 6, 17, 22 UTC) MARGINAL

The worst-performing 4 hours by P/L are: 05:00 UTC (-$2,714), 06:00 UTC (-$2,192), 23:00 UTC (-$1,342), and 07:00 UTC (-$570). Excluding these hours removes 48,561 fills and $150,059 of capital.

Result: +$19,410 P/L on $837,274 deployed = +2.32% ROI, versus baseline +1.92%. The improvement is +$458 in absolute P/L and +0.40 percentage points in ROI.

The marginal improvement is real but minimal. The worst-hour P/L swings (-$2,714 at hour 5) reflect individual market outcomes during those hours, not systematic structural weakness in those time windows. The spread-capture strategy does not have a strong time-of-day edge precisely because the spread opportunity exists whenever the market prices both sides at a combined cost below $1.00 - and that happens around the clock.

Practically: if you are resource-constrained and need to schedule downtime, take it at 05:00-07:00 UTC. You will lose 5.1% of fills but add 0.4 percentage points to ROI. The opportunity cost is small.

5. Combined filter (dom ≥ 2x + exclude worst hours) DESTRUCTIVE

Stacking the hour exclusion on top of the high-conviction dominance filter collapses the sample to 84,843 fills and $154,959 deployed for +$2,296 P/L at +1.48% ROI - worse than both the baseline (1.92%) and the dominance filter alone (12.51%).

The combined filter underperforms the dominance filter alone because the hour exclusion removes fills from the dominance-filtered set that were actually profitable. The stacking assumption - that two individually-positive filters combine multiplicatively - fails because the hour exclusion damages the high-conviction fill set more than it helps it.

Do not stack the hour filter on top of the dominance filter. Run the dominance filter alone.

What filters would genuinely help

The standard PR&R battery misses the actual exploitable dimensions of this strategy. Two genuine refinements that would require additional data:

The paired cost threshold filter would be the single most powerful improvement available if implementable. Currently the median paired cost is $0.9999 (essentially fair value), while the mean is $0.9450. The wide dispersion means many pairs have cost > $1.00 (which should never be entered) and many have cost < $0.90 (which are the best opportunities). Filtering to keep only pairs below $0.95 would dramatically concentrate capital on the highest-spread opportunities.

Bottom line for replication

Three concrete recommendations:

1. Run the high-conviction dominance filter. Copy only fills where the dominant side is 2x+ versus the counter-leg. This is where the real P/L lives: +$90,538 at +12.51% ROI versus +$18,952 at +1.92% unfiltered.

2. Do not apply the $0.30-$0.70 price filter. It destroys the strategy by removing the $0.90+ winning-side fills that carry 56% of capital and most of the settlement payouts.

3. If you must schedule downtime, use 05:00-07:00 UTC. The -$4,906 combined P/L in those 2 hours (05:00 -$2,714 and 06:00 -$2,192) is the only systematic negative window in the 24-hour cycle. Avoiding those hours adds approximately $458 to net P/L with minimal fill-volume sacrifice.

The strategy is already well-optimized at the market selection and sizing level. The standard filter battery reveals one transformational improvement (dominance filter) and confirms everything else the bot is already doing implicitly.

Source wallet: 0x912a58103662ebe2e30328a305bc33131eca0f92 Strategy: Both-sides spread capture with conviction-scaled directional skew on 15-minute crypto Up/Down markets Reference book: $987,333 BUY notional, 284,661 fills, 28 days, +$18,952 net P/L (+1.92% ROI) High-conviction filtered view: $723,554 on dominant legs, +$90,538 P/L (+12.51% ROI)

One-paragraph operator brief

Build a Polymarket bot that enters both sides of every active 15-minute BTC, ETH, and SOL Up/Down window. On each market, buy the Up side and the Down side simultaneously, walking the orderbook on both sides to accumulate shares below your target paired cost of $0.97. Apply a directional signal to size one side 2-5x heavier than the other when you have conviction about the within-window price direction (momentum, orderbook imbalance, or spot price trajectory). Hold everything to settlement - never sell. Run 24/7 with maximum uptime. The strategy earns through two channels: guaranteed spread on pairs where combined cost is below $1.00, and directional P/L where the signal correctly identifies the heavier side. Expect approximately $677/day at the reference scale ($987K turnover/month), with the high-conviction filtered variant generating roughly $3,234/day in P/L on $25,843/day of dominant-leg capital.

1. Market selection

Why 15-minute only: The 15-minute window provides enough within-window price movement for a directional signal to develop while still offering mispriced both-sides pairs at market open. 5-minute windows resolve too quickly for the directional accumulation pattern visible in the CSV (the bot builds positions over 2-14 minutes). 1-hour windows have too much competition from professional market makers who have time to reprice efficiently.

Asset priority: BTC is the highest-volume asset. ETH windows show higher P/L per dollar deployed in the best-performing markets (best ETH market: +$594 on $77 deployed = +774%). SOL appears in the reference book but at lower volume. Priority order: BTC first (capacity), ETH second (efficiency), SOL third.

2. Entry logic

def should_enter_market(market, clob_state):
    # Asset/duration whitelist
    if market.asset not in ("BTC", "ETH", "SOL"):
        return None
    if market.duration_minutes != 15:
        return None

    # Time-within-window gate
    elapsed_seconds = (now() - market.start_time).seconds
    if elapsed_seconds > 840:  # 14 minutes in: too late to pair
        return None

    # Paired cost gate
    up_ask = clob_state.up_side.best_ask
    down_ask = clob_state.down_side.best_ask
    if up_ask is None or down_ask is None:
        return None
    paired_cost = up_ask + down_ask
    if paired_cost > 0.97:
        return None  # spread too thin; wait or skip

    # Directional signal
    direction, confidence = compute_directional_signal(market, clob_state)
    return {"up_ask": up_ask, "down_ask": down_ask,
            "direction": direction, "confidence": confidence,
            "paired_cost": paired_cost}

def compute_directional_signal(market, clob_state):
    # Implementation options (any one or combination):
    # 1. Spot price momentum: compare current BTC/ETH spot to window open price
    # 2. Orderbook depth asymmetry: Up depth vs Down depth on CLOB
    # 3. Mid-price drift: how much has the Up mid moved since window open
    # Returns: ("Up" or "Down", confidence 0.0-1.0)
    ...

Entry pattern from CSV: The bot does not fire once per market. It fires repeatedly over the 15-minute window, building position in 2-second to 2-minute intervals. In the reference CSV, the April 7 BTC 8PM-8:15PM market shows 30+ fills spanning 8 minutes (00:00:57 to 00:14:03 UTC). The bot is continuously polling and filling whenever the ask queue offers shares at acceptable prices.

3. Directional sizing logic (the conviction engine)

The directional sizing is the single most important implementation detail. It is what distinguishes Ratue from a simple both-sides bot and what drives the 85.2% win rate at 3.0x+ dominance.

def compute_position_sizes(entry_data, available_capital):
    direction = entry_data["direction"]
    confidence = entry_data["confidence"]  # 0.0 to 1.0

    # Base allocation: equal on both sides
    base_size = min(available_capital * 0.005, 50.0)  # $0.50-$50 base

    # Dominance multiplier based on confidence
    if confidence < 0.25:
        dom_multiplier = 1.0   # no skew, equal sizing
    elif confidence < 0.50:
        dom_multiplier = 1.5   # 1.5x on favored side
    elif confidence < 0.75:
        dom_multiplier = 2.5   # 2.5x - enters the 2-3x bucket
    else:
        dom_multiplier = 4.0   # 4x+ - enters the 3x+ high-conviction bucket

    favored_size = base_size * dom_multiplier
    counter_size = base_size  # counter-leg stays at base

    # Budget constraint
    total = favored_size + counter_size
    if total > available_capital:
        scale = available_capital / total
        favored_size *= scale
        counter_size *= scale

    return {
        direction: favored_size,
        opposite(direction): counter_size
    }

4. Paired cost management

The paired cost discipline is the spread-capture foundation of the strategy. Every fill decision must maintain awareness of the cumulative paired cost on each market.

class MarketPosition:
    def __init__(self, market_id):
        self.up_usdc = 0.0
        self.up_shares = 0.0
        self.down_usdc = 0.0
        self.down_shares = 0.0

    def paired_cost(self):
        if self.up_shares == 0 or self.down_shares == 0:
            return None  # unpaired, cannot compute
        up_vwap = self.up_usdc / self.up_shares
        down_vwap = self.down_usdc / self.down_shares
        return up_vwap + down_vwap

    def expected_spread_pnl(self):
        if self.paired_cost() is None or self.paired_cost() >= 1.00:
            return 0.0
        paired_shares = min(self.up_shares, self.down_shares)
        return paired_shares * (1.00 - self.paired_cost())

    def should_continue_filling(self, new_paired_cost):
        # Accept new fills if they do not push cumulative paired cost above 0.97
        if self.paired_cost() is None:
            return new_paired_cost < 0.97
        # Weight new fill against existing position
        return True  # simplification: bot continues filling within window

Paired cost targets:

The reference wallet's mean paired cost of $0.9450 sits in the $0.90-$0.97 range, confirming this is the typical entry zone.

5. Sizing model and bankroll

Fixed vs scaled sizing: The reference bot uses conviction-scaled sizing (power-law distribution, top 5% of fills carry 48.7% of capital). Implement the dominance-ratio sizing model above; do not use fixed clips.

Bankroll requirements at various scales:

The capital requirement is high because the strategy involves buying both sides, which means each market consumes 2x the per-side notional. With a $50,000 liquid balance and positions cycling every 15 minutes, the bot can maintain 3-5 concurrent active markets at $2,000-$5,000 average notional each.

Gini / concentration: The reference book has a Gini of approximately 0.70 (top 5% of fills carry 48.7% of capital). Expect and design for high concentration in high-conviction fills.

6. Exit logic

There is no exit logic. All positions hold to settlement.

This is a critical design choice. The reference wallet has 0 SELL trades across 284,661 fills. Every share purchased rides to $1.00 (win) or $0.00 (loss) at window resolution.

This design works because: 1. The paired structure guarantees that one side wins and one side loses. The net is the spread, regardless of direction. 2. Selling before resolution would require finding a buyer willing to pay more than the current market price, which is uncertain and unnecessary given the guaranteed settlement. 3. The 15-minute window is short enough that holding to settlement is always the fastest exit anyway.

Do not add a SELL leg to this strategy. It would add execution complexity, latency risk, and potential slippage that the guaranteed settlement eliminates.

def on_window_close(market, positions):
    # No action required - settlement is automatic
    # Polymarket settles all positions at $1.00 (winner) or $0.00 (loser)
    # Log the outcome for diagnostics
    log_settlement(market, positions)
    del active_positions[market.id]

7. Hour scheduling and operational cadence

Unlike SirMartingale, Ratue runs 24/7. The reference book has fills in every single UTC hour. The variance in hourly P/L (-$2,714 at 05:00 to +$3,486 at 03:00) reflects individual market outcomes, not systematic time-of-day edge.

Recommended scheduling:

15-minute window management: Every 15 minutes, a new set of markets opens (btc-updown-15m-TIMESTAMP, eth-updown-15m-TIMESTAMP, sol-updown-15m-TIMESTAMP). The bot must detect new market creation, enter both sides within the first 2-3 minutes, and continue accumulating until 14 minutes elapsed.

Concurrent active markets: at any given time, the bot has 1-3 active markets open (the current 15-minute window, sometimes overlapping with a market that opened 30 seconds ago). Position management across concurrent markets is the primary operational complexity.

8. Operational requirements

9. Risk management

The worst-case scenario: A run of markets where the dominant side is wrong at high dominance ratio. In the reference book, the 3.0x+ bucket had 85.2% dom WR across 4,249 markets. If this dropped to 50% (random) for a sustained period, the directional P/L would disappear and the spread P/L alone would determine returns. Spread P/L at $0.9450 mean paired cost would generate approximately $0.055 per paired share - positive but thin. The strategy would earn less, not bleed out.

Per-market loss cap: Set a maximum USDC deployment per individual market of $1,500 (based on worst observed loss of -$684). This prevents any single mis-priced market from exceeding 3% of a $50K bankroll.

10. Diagnostic checklist - is the bot still working?

Run weekly:

What this playbook deliberately does not include

No exit / SELL logic. The reference wallet has 0 SELL trades. Adding a SELL leg introduces execution risk, latency dependency, and partial-fill risk that the settlement guarantee eliminates. Do not add it.

No multi-category expansion. The strategy is calibrated specifically to 15-minute crypto Up/Down markets. Applying the same paired-cost logic to sports, politics, or hourly windows would require rebuilding the directional signal from scratch and validating the paired cost opportunity in those markets independently.

No Kelly sizing. Kelly would concentrate too much capital on high-conviction markets and create ruinous drawdown potential when the directional signal fails. The reference book's power-law sizing (4x maximum dominance) is already aggressive. Do not add further concentration.

No copy-trading hook. This strategy generates its own directional signal from within-window price action. It does not copy another wallet, and there is no wallet to copy from - the fills are on markets that expire in 15 minutes, which is too short for copy-following to be viable.

No hedging or correlation management. Ratue runs BTC, ETH, and SOL simultaneously in the same 15-minute windows. Those three assets are correlated. A simultaneous adverse move in all three generates correlated losses. The reference book shows this is acceptable given the spread-capture floor return, but a risk-averse replicator might run only one asset at a time. Expected P/L impact: roughly 1/3rd at BTC-only, consistent with BTC's share of total volume.

The entire strategy rests on two pillars: finding pairs below $0.97 and sizing the favored side 2-5x heavier when you have a directional view. Everything else is implementation detail. Build those two pillars correctly and the rest follows mechanically.