Finance LLM Leaderboard — AI Accuracy for Financial Tasks

Why Standard AI Benchmarks Don't Measure Financial Accuracy

MMLU (Massive Multitask Language Understanding) tests general knowledge across 57 subjects, including some finance. HumanEval tests code generation. HELM tests reasoning. These are the benchmarks you'll see cited when AI vendors claim their model is "the best." None of them measure what finance professionals actually need to know: does the model compute the correct Black-Scholes price? Does it apply the Hamada equation correctly when relevering beta for a target capital structure? Does it produce an ASC 842 lease liability that matches the FASB standard?

The difference matters enormously. A DCF discount rate wrong by 80 basis points changes the implied share price by 8–12%. A VaR calculation off by 15% is a risk management failure. A misapplied MACRS depreciation table is an IRS penalty risk. "Looks plausible" is not a pass criterion that auditors, actuaries, or regulators accept. It should not be the standard for AI finance tools either.

What the ClaudeFinanceLab Eval Framework Tests

The ClaudeFinanceLab financial accuracy benchmark uses deterministic, analytically computed ground truth across six evaluation suites:

• Black-Scholes and Options Greeks — call and put pricing (closed-form), delta (N(d₁)), gamma (N'(d₁)/Sσ√T), vega, theta, rho. Pass criterion: within 0.001 of the closed-form value for pricing, within 0.0001 for Greeks.
• WACC and Hamada Equation — weighted average cost of capital, cost of equity via CAPM, Hamada relevering/unlevering beta, country risk premium adjustments. Pass criterion: within 0.01 percentage points.
• MACRS Depreciation — all asset classes (3-, 5-, 7-, 10-, 15-, 20-year property), half-year and mid-quarter conventions, switching rules. Ground truth: IRS Rev. Proc. 87-57 tables. Pass criterion: exact match to IRS table.
• ASC 842 Lease Accounting — operating and finance lease liability (present value of future lease payments at IBR), right-of-use asset, amortization schedule. Pass criterion: within $1 of FASB-compliant computation.
• DCF Valuation — free cash flow projections, terminal value (Gordon growth model), enterprise value, equity bridge. Pass criterion: within 0.1% of analytically computed value.
• CAPM and Capital Structure — expected returns, cost of equity, Modigliani-Miller adjustments, levered/unlevered beta. Pass criterion: within 0.01 percentage points.

Finance Accuracy: Claude + MCP vs. Raw LLM Reasoning

The Claude + MCP column below is our own published Track 1 benchmark — deterministic, reproducible, live on the Benchmark page right now. Head-to-head comparisons against GPT-4o, Gemini, and Claude without specialized tools are in progress; we publish results only once they're actually run through the same methodology, not estimated.

"Claude + MCP" pass rates are live, reproducible results from our public Track 1 benchmark — deterministic ground truth, re-run on every deployment. The "Claude (no MCP)", GPT-4o, and Gemini columns are not yet populated: we publish a comparison only once it has actually been run through the same methodology, not before. Check back, or get in touch if you'd like early access to that data.

Why MACRS and ASC 842 Have Such Low LLM Pass Rates

MACRS depreciation is table-driven — the correct answer requires looking up the exact rate from IRS Rev. Proc. 87-57, not computing it from first principles. A raw LLM asked for a MACRS rate is generating a plausible-looking percentage from pattern-matching, not retrieving one from a memorized table — the kind of task where hallucination risk is highest and hardest to catch by eye. That's exactly the class of error a deterministic, table-driven MCP tool eliminates by construction: it looks up the rate, it doesn't guess it.

ASC 842 failures follow a different pattern: most LLMs correctly understand the conceptual framework but make errors in the present value calculation — wrong incremental borrowing rate, misapplied payment timing (beginning vs. end of period), incorrect treatment of initial direct costs, or ignoring the practical expedient elections. Even a 3% error in the lease liability present value is material for large lease portfolios.

The ClaudeFinanceLab MCP tools solve this by calling the calculation function directly with the correct parameters — the Black-Scholes MCP tool doesn't generate a Black-Scholes price, it computes one from the Scholes-Merton formula with defined precision. The test result is a pass/fail against that formula, not against what another LLM thinks the answer is.

What This Means for Finance AI Procurement

When evaluating AI tools for financial use, the right question is not "what benchmark score does this model get on MMLU?" It's: "has this tool been tested against the specific calculations my team relies on, with documented ground truth and published pass rates?"

Use these prompts to run your own accuracy spot-check on any AI finance tool:

• "Price a European call option with these inputs: S=100, K=105, T=0.5 years, risk-free rate=5%, volatility=20%. Show your calculation of d₁, d₂, N(d₁), N(d₂), and the final call price. The analytically correct answer is $5.573."
• "Calculate WACC for a company with: market cap $500M, debt $200M, cost of equity 12% (CAPM: risk-free 4.5%, beta 1.1, ERP 6.8%), pre-tax cost of debt 6.5%, tax rate 25%. Show each component. The correct answer: E/(D+E) = 0.714, D/(D+E) = 0.286, after-tax cost of debt = 4.875%, WACC = 12% × 0.714 + 4.875% × 0.286 = 9.955%."
• "What is the MACRS Year 1 depreciation rate for a 5-year property class under the half-year convention? The IRS table answer is exactly 20.00%. If the tool gives any other number, it is wrong."
• "A company has an operating lease with 5 annual payments of $100,000 starting one year from now. The incremental borrowing rate is 6%. What is the opening right-of-use asset and lease liability? The correct answer (PV of annuity: 100,000 × [1-(1.06)^(-5)]/0.06) = $421,236. If the tool gives a different number, it has made an arithmetic error."

The Full Eval Framework

The complete ClaudeFinanceLab financial accuracy benchmark — test IDs, inputs, expected outputs, actual outputs, and methodology — is publicly available at claudefinancelab.com/evals/. All tests are deterministic and reproducible. Results are re-run on every deployment of the platform.

Related Reading

• ClaudeFinanceLab Financial Accuracy Evals — live benchmark results
• Finance AI Model Validation — SR 11-7, SR 26-2 & EU AI Act
• AI Hallucination in Finance — Detection & Prevention
• AI for Model Risk Management — SR 11-7, Validation & MRM
• Best AI Financial Tools — Accuracy Rankings 2026