Benchmarks & Evidence

Proof, not adjectives. Price the whole book before the open.

PRISM is the tax-aware optimization core that scales to an entire personalized book — deterministically, auditably, with no per-seat solver license. Every figure below is measured on real US-equity data (March–April 2026), stated with its regime, with losing cases kept in.

10.2×
Fleet throughput vs a commercial CPU baseline (144 vs 14 accounts/sec/core)
11.6min
To price 100,000 personalized accounts on one core (vs ~118 min)
100,004
Real assets backtested at ~1.8s/rebalance · Sharpe 0.737
$238k
Full harvestable tax budget captured on a $5M, 192-name book
Evidence integrity. Every number traces to one recorded benchmark run on real US-equity data. Comparators are labeled generically. We keep losing cases in — no cherry-picking, no stitched numbers, no parity we didn't measure.
01 · Scale incumbents can't reach

The workflow exact solvers physically can't run.

Nobody runs an exact solver across a 100,000-account book before the open — it times out long before. PRISM does, deterministically, and stays fast at every universe size.

Full-market backtest
100,004
Real assets across 47+ exchanges, solved at ~1.8 seconds per rebalance. Walk-forward Sharpe 0.737, max drawdown −4.22%.
Exact baseline ceiling
~5,000
Where an exact commercial solver times out (~33 s). PRISM prices the full book past that ceiling with no timeout.
Fastest measured
190–40k
PRISM is the fastest engine measured at every universe size across this range — not a single cherry-picked point.
02 · Fleet throughput

~10× more accounts per core — the margin lever.

Personalized accounts grow faster than AUM. Throughput per core decides how many accounts an ops team can run, and whether the nightly batch clears on time.

PRISMaccounts / sec / core
144
Commercial CPU Baseline 1accounts / sec / core
14
100,000 accounts · PRISM
11.6 min
Priced on a single core — and it parallelizes linearly across cores.
100,000 accounts · baseline
118 min
The same book on the commercial CPU baseline, single core.
Determinism
bit-exact
Repeated runs return content-hashed, identical outputs — no solver timeouts, no flaky reruns.
03 · Tax alpha

The only engine still harvesting six figures at scale.

As beta commoditizes, after-tax outcome is how direct-indexing and SMA desks compete. This is where PRISM is measured strongest — and where simpler approaches quietly leave money on the table.

ApproachTax budget captured · $5M, 192 namesSix-figure alpha at 5,000+ names?After-tax win rate vs PRISM
PRISM~$238k (full budget)Yes
Open-source solver baseline~$3k at scaleNoPRISM wins ~88%
Rules-based tax-loss harvestingpartialNoPRISM wins ~95%

Tax figures are measured on real-calibrated books; the robust real metric (tax captured) is measured directly. Your numbers are computed live on your data during the pilot.

04 · Quality, honestly

Near-exact — and we name where we don't win.

A sophisticated buyer trusts the vendor who states their non-fit first. So here it is, both sides.

Where PRISM wins

  • Matches the exact reference optimum to ~1e-7.
  • Competitive with an exact commercial baseline on after-tax outcome: matches it in 58–63% of regimes, wins outright in ~33%.
  • Runs the fleet-scale and full-universe workflows where exact solvers can't run at all.
  • Deterministic, reproducible, auditable — every trade traces to your constraints and tax logic.

Where PRISM does not

  • It does not beat the exact single-account optimum every time — no heuristic does.
  • It is not a bull-market beta play; diversification and tax discipline have a cost.
  • The quantum lane is a research track that adds zero tax alpha today — the production results are classical and measured.
05 · The measured curves

Every headline number, charted.

The same evidence behind the claims above, plotted straight from the recorded runs — factor-constrained QP on real US-equity universes, March–April 2026, p50 of five timed runs. Lower is better on solve time; speedup is versus Gurobi 13.0.1 on CPU.

Figure 1 · Solve time vs universe size — factor-constrained QP
1000 750 500 250 0 solve time p50 (ms) 5,000 7,000 9,000 universe size (assets) HiGHS · 994.8 ms Gurobi · 541.4 ms PRISM · 303.6 ms
PRISM (GPU) Gurobi 13.0.1 (CPU) HiGHS (CPU)

PRISM grows sub-linearly while the CPU baselines climb faster: at N=9,000 PRISM finishes in 303.6 ms versus 541.4 ms (Gurobi) and 994.8 ms (HiGHS), all at sub-ppm quality gaps.

Figure 2 · Speedup by transition workflow — PRISM (GPU) vs Gurobi 13.0.1 (CPU), N=5,000
Transition Rebalancegap < 0.004%
124.85×
Market Impactgap < 0.01%
25.86×
Tax-Aware Transitiongap < 0.01%
21.32×
Crisis Resiliencegap < 0.001%
20.92×

Every transition workflow clears a 20× floor against an exact commercial baseline, with every quality gap held under 0.01%. The rebalance lane — the most common intraday operation — reaches 124.85×.

Solver Runtime p50 (ms) Gap vs Gurobi Stability (CV)
PRISM (GPU)138.7< 0.00001%11.7%
Gurobi 13.0.1 (CPU)208.1reference19.2%
HiGHS (CPU)466.9< 0.00001%10.0%
Clarabel (CPU)768.6< 0.00001%12.0%
cuOpt (GPU)870.00.0014%8.0%
OSQP (CPU)11,311< 0.00001%10.0%

Figure 3 · Structured QP at 5,000 real assets. PRISM is the fastest exact-quality solver in the field — 1.5× under Gurobi and 3.4× under HiGHS — while staying within a tight stability band. We keep the honest qualifier in: cuOpt posts a lower run-to-run CV, and Gurobi remains the correctness reference every other engine is measured against.

Figure 4 · Fleet throughput — personalized accounts priced per second, per core
PRISM (GPU)~11.6 min for 100,000 accounts
144 acct/s
Commercial CPU baseline~118 min for 100,000 accounts
14 acct/s

~10.2× per-core throughput, and PRISM parallelizes linearly across cores. The whole personalized book is priced inside the overnight window, before the open — the workflow the fleet is actually billed on.

Figure 5 · The advantage holds at the ceiling — speedup vs Gurobi at N=75,257 real assets
Market Impact
12.28×
Tax-Aware Transition
5.60×
Rebalance
4.42×
Crisis Resilience
2.27×

Speedup peaks at 124.85× at N=5,000 and settles into a durable multiple as the universe grows to the full real-data ceiling. The institutional question at scale is persistence, not a monotonic climb: every transition workflow stays multiples faster — in a regime where exact solvers no longer finish at all.

SolverQP SpeedReplayTransitionQualityStabilityMax Scale
PRISM (GPU)10101010910
Gurobi (CPU)7631078
HiGHS (CPU)650975
cuOpt (GPU)521265
OSQP (CPU)221354
Clarabel (CPU)410934

Figure 6 · Competitive overview — six solvers across six dimensions, scored 0–10 on real-data benchmarks (March 2026); higher is better. PRISM leads on speed, transition workflows, and scale; Gurobi and Clarabel match it on single-problem quality; cuOpt edges run-to-run stability. No solver wins every axis — which is the point of showing all six.

06 · Trust primitives

Built to pass compliance, not just benchmarks.

In this buyer set, trust is the product and compliance is the gate. PRISM ships the de-riskers up front.

Deterministic & reproducible

Content-hashed, bit-reproducible outputs — re-derivable for any audit or exam date.

Tax-rule correctness

Wash-sale handling and lot-level tax accounting, validated by an internal test suite — 18 tests pass.

Validate against ground truth

An exact-reference comparator ships so you can check PRISM against the exact optimum on your own data.

Transparent, not a black box

Every trade traces to explicit constraints and tax logic — no opaque ML in the trade path.

07 · How we benchmark

Matched workloads. Honest qualifiers. Losses kept in.

A benchmark is useful only when it tells you whether an engine can clear your workload with acceptable quality, latency, and reliability — not when it collapses to one headline number.

Real data
Core results run on real US-equity data and a real 100,004-asset benchmark — not toy problems.
Matched comparison
Same universe, constraints, costs, and tax rules across every engine. Quality gap reported alongside runtime.
Reproducible
Fixed seeds, recorded runs, low run-to-run variance. Every figure ties to an artifact.
On comparators: baselines are labeled generically — Commercial CPU Baseline 1 / 2 (exact and fast-lane commercial solvers), open-source solver baselines, and rules-based tax-loss harvesting. We compare on outcomes (after-tax wealth, throughput, scale, feasibility), state the regime, and include the cases where a baseline wins. PRISM's results describe outcomes; its methods are proprietary. Read more: how to read optimization benchmarks without being misled →

The proof that counts is on your data.

A 30-day, buyer-owned matched-workload pilot: PRISM vs your current stack, on your universe, constraints, costs, and tax rules — a pass/fail metric you set before we start. You get a full results pack (every account, losses shown), deterministic audit logs, and an ROI computed with your real numbers.

Request a matched-workload pilot →