Pillar A · State as of 2026-08-07

Is quantum computing useful for financial risk management today?

No — as of August 2026, no quantum computer has beaten a strong classical Monte Carlo engine on a real risk workload (VaR, CVaR, greeks) on the same instance. The quadratic speedup from quantum amplitude estimation is real mathematics, proven on paper since 2015. But published resource estimates put the crossover at ~8,000 logical qubits running at tens of MHz of logical clock — hardware that does not exist — and a quadratic speedup is exactly the kind that error-correction overhead eats first. Every hardware demo so far is component-scale: 5 qubits for VaR, 3 for option pricing, 64 for data loading.
→ Leer en español
State as of: 2026-08-07

Status as of: August 2026.

Financial risk is the class where quantum computing has its cleanest theoretical story and one of its emptiest scoreboards. The story: risk engines run Monte Carlo, Monte Carlo converges as 1/√M, and a quantum algorithm — amplitude estimation — converges as 1/M. A provable quadratic speedup on the workhorse computation of every bank. The scoreboard: as of August 2026, nobody has run that speedup end-to-end against a strong classical Monte Carlo engine on the same risk instance and won. Not on Value-at-Risk, not on CVaR, not on greeks. This post lays out both halves with sources, because the gap between them is the actual state of the art.

What does the theory actually promise?

The chain has three links, all real. Brassard, Høyer, Mosca and Tapp formalized quantum amplitude estimation in 2002: estimate a probability amplitude to error ε with O(1/ε) oracle calls, where classical sampling needs O(1/ε²). Montanaro (2015) turned that into a general quantum speedup for Monte Carlo: quadratically fewer samples for the same error bar. Woerner and Egger (IBM, npj Quantum Information, Feb 2019) mapped it to the numbers a risk desk actually reports — VaR and CVaR — and demonstrated the circuit on hardware.

The catch is not in the math. It is in three places the math does not cover: the constant factors of fault tolerance, the cost of loading market data into superposition, and the strength of the classical opponent. A quadratic speedup halves the exponent; it does not survive a 10⁶× constant-factor overhead until problem sizes get enormous. That is not a rhetorical point — it is the published conclusion of the people building the machines, as the crossover section below shows.

THE EVIDENCE LADDER — QUANTUM RISK, AUG 2026Theory: quadratic speedup (QAE)PROVEN — Brassard 2002 · Montanaro 2015Hardware demos: VaR, option pricingTOY SCALE — 5 qubits (2019) · 3 qubits (2020)Data loading: heavy-tail encodingCOMPONENT — 25–64 qubits, IBM HW (2025)Risk advantage vs strong classical MCNOBODY — no published end-to-end win, any instanceTeal = done · Gold = partial or absent

What has actually been measured on hardware?

Every hardware result in this class is a component test, not a race. The table is short because the honest table is short.

Result Scale What it shows What it does not show Source (date)
Quadratic speedup for Monte Carlo theory 1/M vs 1/√M convergence constant factors, data loading Montanaro, Proc. R. Soc. A (2015)
VaR/CVaR via amplitude estimation 5 qubits, T-bill toy circuit works on real HW any advantage — instance is trivial Woerner & Egger, npj QI (Feb 2019)
Option pricing on hardware 3 qubits end-to-end pricing circuit scale; error bars beat nothing Stamatopoulos et al., Quantum 4, 291 (Jul 2020)
Heavy-tail distribution loading 25–64 qubits (IBM Eagle/Osprey) MPS shallow circuits ease the data bottleneck a risk number; it is an encoder, not a solver HSBC + Haiqu, via FinTech Magazine (2026)
Advantage threshold, derivative pricing resource estimate what the crossover costs that HW exists — it does not Chakrabarti et al., Quantum 5, 463 (Jun 2021)

Note what is absent: any row where a quantum device produced a VaR, CVaR or greek faster or cheaper than a tuned classical Monte Carlo on the same instance. Classical Monte Carlo is also embarrassingly parallel — it scales across GPU farms with near-zero coordination cost — so the classical baseline is not standing still while the quantum side matures. Our post on why a weak classical baseline ruins a quantum benchmark applies to this class in full.

Where would the crossover be?

This is the rare class where the crossover has been costed by the optimists, in public, with numbers. Chakrabarti et al. (Goldman Sachs + IBM, Quantum, Jun 2021) estimated what it takes for quantum derivative pricing to beat classical: ~8,000 logical qubits and a T-depth of 54 million, executed in about a second — a logical clock rate on the order of 50 MHz. Stamatopoulos et al. (Quantum, Jul 2022) improved the greeks pipeline with quantum gradient methods and cut the requirement to ~7 MHz logical clock — still purely a paper result, derived from simulation, assuming fault-tolerant hardware.

For calibration: no fault-tolerant machine with thousands of logical qubits exists at any clock rate; vendor roadmaps place ~200 logical qubits in 2029 and ~2,000 in 2033 (we read those roadmaps here — a roadmap is a goal, not a schedule). And Babbush et al. (Google, PRX Quantum, 2021) made the structural point that applies to this whole class: quadratic speedups do not deliver net advantage on early fault-tolerant machines — the error-correction overhead consumes them. Risk's speedup is exactly quadratic. That is why the same analysis that motivates quantum risk research also dates it beyond the first FT generations, unless the constants fall dramatically.

CROSSOVER ASK vs 2026 REALITYASK (Chakrabarti 2021 · Stamatopoulos 2022)~8,000 logical qubits · T-depth 5.4×10⁷ · ~50 MHz(~7 MHz for greeks via gradient methods — paper only)EXISTS (measured, this class)5-qubit VaR demo · 3-qubit pricing · 0 logicalqubits at the required scale on any machineBars not to scale — the true gap does not fit on a chart.

Didn't HSBC already claim a quantum advantage in finance?

A claim exists; it is not a risk result, and it deserves a careful reading. On 25 Sep 2025, HSBC and IBM announced a "world-first" in quantum-enabled algorithmic trading: up to ~34% improvement in predicting bond RFQ fill probability, using IBM Heron processors to transform features for a classical ML model. Three things keep this off the risk scoreboard. First, the task is ML feature generation for trade-fill prediction — not Monte Carlo risk, not amplitude estimation; no quadratic-speedup mechanism is involved. Second, the team's own paper notes that hardware noise contributes to the gain, and that the noiseless simulation of the same quantum computation does not show the advantage — which is the opposite of what a computational speedup looks like. Third, independent replication has not happened; Aaronson's public critique (Sep 2025) details why noise-as-regularizer effects need classical noise baselines before "advantage" is the right word. None of this is an accusation — HSBC's release describes the experiment accurately. It is a reminder that in this class, the measured wins so far are not the kind the theory promises. Our checklist for reading advantage claims applies.

What would change the verdict?

Three observable events, any of which we would report as news: (1) a fault-tolerant machine with thousands of logical qubits at MHz-scale logical clock rates — watch actual logical-qubit counts, not physical; (2) an end-to-end amplitude-estimation risk computation beating a tuned classical Monte Carlo on the same instance, same accuracy, wall-clock and dollars reported — cost per answer, not cost per shot; (3) an algorithmic improvement that upgrades the quadratic speedup to something error correction cannot eat. Until one of those lands, the honest verdict for risk desks is: keep your GPUs, watch the crossover math, and treat every "quantum finance" headline as a claim to be parsed, not a result to be bought.

What we know / what we don't know

What we know. The quadratic speedup for Monte Carlo is proven mathematics (2002/2015). VaR/CVaR circuits run on real hardware at toy scale (2019). The crossover has been costed by proponents at ~8,000 logical qubits and tens of MHz of logical clock (2021), improvable to ~7 MHz for greeks on paper (2022). Quadratic speedups are the first casualty of error-correction overhead (2021). No end-to-end quantum win on a real risk workload has been published as of August 2026.

What we don't know. Whether algorithmic constants will fall faster than hardware roadmaps assume — the field has surprised before (dequantization surprised in the other direction). Whether the data-loading bottleneck (encoding market distributions into superposition) has a general solution; MPS encoders are progress on one piece, not the whole. Whether noise-enhanced ML effects like the HSBC result replicate under classical noise baselines — nobody has published that test. And, disclosed plainly: Rosetta has no sealed runs in this class. Our measured evidence (the V-0012 series) is in small-instance portfolio optimization — a different class (that verdict is here) — and we claim nothing measured about risk workloads. A dedicated risk bench is on our backlog; until it runs, our verdict here rests entirely on the published record above.

Sources

Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.

Sources:
· Montanaro, Quantum speedup of Monte Carlo methods, Proc. R. Soc. A 471, 20150301 (2015)
· Brassard, Høyer, Mosca & Tapp, Quantum amplitude amplification and estimation (2002)
· Woerner & Egger, Quantum risk analysis, npj Quantum Information 5, 15 (Feb 2019)
· Stamatopoulos et al., Option pricing using quantum computers, Quantum 4, 291 (Jul 2020)
· Chakrabarti et al., A threshold for quantum advantage in derivative pricing, Quantum 5, 463 (Jun 2021)
· Babbush et al., Focus beyond quadratic speedups for error-corrected quantum advantage, PRX Quantum 2, 010103 (2021)
· Stamatopoulos, Mazzola, Woerner & Zeng, Towards quantum advantage in financial market risk, Quantum 6, 770 (Jul 2022)
· HSBC media release, quantum-enabled algorithmic trading with IBM (25 Sep 2025)
· Aaronson, on the HSBC–IBM claim (Sep 2025)
· FinTech Magazine, HSBC & Haiqu heavy-tail encoding on IBM hardware (2026)