Execution Analytics CONFIDENTIAL
19,439 orders · USD 22.6bn notional · Jan 2025 – Apr 2026 · LATAM excluded from all analysis
Overview
Key findings
Statistical analysis of IS and CUSTOM execution across six core brokers. All figures winsorised P1/P99. Benchmark: is_at_touch_bps.
IS STRATEGY
ADV cliff at 3–5%
−5.9 bps
IS cost stays negative to 3% ADV. Sweet spot at 2–3% ADV. Cost turns positive and material above 5% ADV (+12.6 bps).
CUSTOM STRATEGY
Earlier cliff at 1–2%
+51.6 bps
CUSTOM breaks immediately at 1–2% ADV. Above 3% ADV mean exceeds +50 bps. Fundamentally different cost dynamics to IS.
VOLATILITY
Universal non-adaptation
+31 bps
Every broker accelerates in high vol — the opposite of optimal IS behaviour. 60+ bps cohort costs +31 bps mean. Route to JPM/CITI above 42 bps annualised.
TIME OF DAY
15:00 UTC peak cost
+21.3 bps
EMEA IS at 15:00–16:00 UTC (US open overlap) costs 16 bps more than the 16:00 window. Avoid high-participation arrivals in this window.
RANK PERSISTENCE
No monthly persistence
12/13
12 of 13 consecutive monthly broker IS rank-pairs show no significant correlation. Static TCA tiering has no statistical validity in this dataset.
APAC FILLS
Long-patience cluster
−20.8 bps
Long-duration APAC IS (>120 min) achieves the best outcomes in the dataset. 100% of high-quality fill proxy events are APAC. JPM leads this cohort.
Dataset
Coverage and strategy mix
Total orders
19,439
post-dedup, ex-LATAM
IS orders
6,776
34.9% of total
CUSTOM orders
4,241
21.8% of total
LATAM excluded
1,148
commission pool only
Notional
$22.6bn
USD, all strategies
Period
15 mo
Jan 2025 – Apr 2026
IS Strategy
ADV participation threshold
At what order size relative to ADV does IS cost turn positive? The 2–3% band is the sweet spot. The cliff is at 3–5%.
IS mean
IS median
CUSTOM mean
CUSTOM median
IS sweet spot: 2–3% ADV (−5.9 bps mean). Cost is negative across the full 0–3% range. The cliff at 3–5% is real — mean flips to +2.5 bps, median to +8.75 bps.
CUSTOM has no sweet spot. Cost turns positive immediately at 1–2% ADV (+7.2 bps). Above 3% ADV the mean exceeds +50 bps. These are structurally different instruments.
CUSTOM has no sweet spot. Cost turns positive immediately at 1–2% ADV (+7.2 bps). Above 3% ADV the mean exceeds +50 bps. These are structurally different instruments.
IS Strategy
Order ADV × interval participation heatmap
For a given order size (columns), at what interval participation rate (rows) does cost become problematic? Key finding: interval participation rate is not the driver — order ADV size is.
strongly beneficial (<−15)
beneficial
neutral / sparse
costly
severely costly
Cost is driven by order ADV size, not pace. The <1% ADV row stays green or neutral across every interval participation rate — even at >25% pace. High interval participation in small orders reflects liquidity-seeking behaviour (dark pool, periodic auction) that aids performance, not market impact.
IS Strategy
Time of day — EMEA vs non-EMEA
Arrival hour UTC vs IS cost. The 15:00–16:00 UTC window (US open overlap into European close) is the most expensive IS window in the dataset.
IS EMEA
IS non-EMEA
15:00–16:00 UTC: +21.3 bps EMEA IS mean. This is 16 bps above the 16:00 window (+5.2 bps). Avoid initiating high-participation EMEA IS orders in this window where possible.
Non-EMEA is consistently negative — APAC and Americas IS orders show beneficial outcomes across most hours, confirming the regional cost asymmetry.
Non-EMEA is consistently negative — APAC and Americas IS orders show beneficial outcomes across most hours, confirming the regional cost asymmetry.
IS Strategy
Volatility regime × ADV participation
Vol regime compounds participation cost. The same order in high vol costs materially more than in low vol at every ADV band.
low <28 bps
normal 28–42
elevated 42–60
high >60 bps
Universal algo design flaw: every broker accelerates in high volatility — the opposite of optimal IS behaviour. Confirmed for all 6 brokers at p<0.01.
Route IS to JPM or CITI above 42 bps annualised vol. Avoid UBS, BARC, BofA in elevated vol conditions.
Route IS to JPM or CITI above 42 bps annualised vol. Avoid UBS, BARC, BofA in elevated vol conditions.
IS Strategy
Broker × ADV participation band
How each broker's IS cost varies by order size. Reveals which brokers maintain performance as participation increases.
Select a broker to see their IS cost profile by ADV participation band.
CUSTOM Strategy
ADV participation threshold
91.3% of CUSTOM orders complete in under 1 minute — these are single-event crosses. Cost dynamics are structurally different from IS.
Instant fills <1min
91.3%
single-event crosses
APAC mean cost
−5.5 bps
vs EMEA +32.7 bps
Cost cliff
1–2% ADV
turns positive immediately
Notional cliff
≥USD 1m
mean jumps to +41 bps
CUSTOM mean
CUSTOM median
CUSTOM Strategy
Spread quality × notional size
Unlike IS, CUSTOM cost is driven primarily by notional size — not spread quality. Tight spread instruments still incur high cost above USD 2m.
Notional size dominates spread quality above USD 1m. Tight spread instruments at USD 2–5m cost +13.5 bps mean — almost as costly as wide spread at the same size. The exception: wide spread / USD 500k–1m (−25.8 bps, n=96) likely reflects strong periodic auction or dark pool fills in mid-cap European instruments.
CUSTOM Strategy
Broker performance
BARC and BofA show inverted rankings vs IS — confirmed across two independent weight windows. A single weight vector is structurally wrong for both strategies.
BARC and UBS median = 0.00 bps — the majority of their CUSTOM orders execute at benchmark. Elevated means are tail events, not systematic failure.
BofA median = +20.4 bps — this is systematic underperformance, not skew. Confirmed p=0.008.
IS-CUSTOM inversion: BARC underperforms IS but leads CUSTOM. BofA leads IS but underperforms CUSTOM. Strategy-conditional bandit instances are required.
BofA median = +20.4 bps — this is systematic underperformance, not skew. Confirmed p=0.008.
IS-CUSTOM inversion: BARC underperforms IS but leads CUSTOM. BofA leads IS but underperforms CUSTOM. Strategy-conditional bandit instances are required.
Recommender
Value proposition vs random algowheel
Monetary estimate of routing intelligence vs a v1 random algowheel with zero routing logic. IS flow only; CUSTOM upside excluded pending W2 spike investigation.
IS annual notional
~$5.5bn
annualised run rate
Exploitable flow
~30%
high-vol + high-par
Conservative recovery
10 bps
50% capture rate
IS annual saving
$1–2m
central estimate
CUSTOM upside
$1–2m
directional — blocked
Combined range
$2–4m
per year at run rate
| Condition | Random cost | Routed cost | Recoverable | Confidence |
|---|---|---|---|---|
| Vol >60 bps IS | +31 bps | ~+8 bps (JPM/CITI) | ~22 bps | High — p<0.01 all brokers |
| High-par IS (>3% ADV) | +26.8 bps | ~+8 bps (patient routing) | ~18 bps | High — rho=0.181, p<0.0001 |
| Fast-illiquid IS | +32.2 bps | +25 bps (patient broker) | ~7 bps | High — p=0.044 |
| Long APAC patience | ~−5 bps (random hit rate) | −20.8 bps (targeted) | ~15 bps | Medium — n=118 |
| CUSTOM IS-inversion | BARC/BofA random mix | Strategy-conditional | ~14 bps | High — two windows confirmed |
Conclusion
Expert assessment
// RECOMMENDER vs RANDOM — VERDICT: B
The recommender is the correct choice, unambiguously. Random allocation is the naive comparator — and the data quantifies the margin it leaves on the table at USD 1–2m per year on IS flow alone at current run-rate notional.
The volatility non-adaptation finding alone justifies the system. Every broker accelerates in high volatility — the opposite of optimal IS behaviour, confirmed for all six brokers at p<0.01. Random allocation routes to UBS or BARC in a high-vol EMEA IS order as often as it routes to JPM or CITI. The recommender routes to JPM or CITI every time that condition is met.
The IS-CUSTOM inversion for BARC and BofA is confirmed across two independent weight windows. A single weight vector is structurally wrong for both strategies simultaneously. Strategy-conditional bandit instances are required architecture, not an optimisation.
Build now: IS bandit with strategy conditioning, LATAM excluded from reward signal, participation rate as primary routing gate, confidence-weighted by broker sample density.
Do not build yet: CUSTOM reward function — W2 spike of +44.8 bps requires investigation with LiquidMetrix before the signal is trustworthy.
Immediate action: UBS is the only significantly elevated IS broker in W2 (p=0.013). TS implied weight 0.6%. Current 18% allocation is not supported by evidence.
Source: latest.parquet, 19,439 orders, Jan 2025–Apr 2026. LATAM (1,148 orders) retained in commission pool, excluded from all performance analysis. Methods: Mann-Whitney U two-sided; Spearman rank correlation; K-means k=5; Thompson Sampling Gaussian conjugate posterior. Winsorisation: global P1/P99. Commission: 3 bps × USD 22.56bn annualised = USD 6.77m pool.