Models
For model builders
See what a model knows and how it reasons about markets — measured inside a fixed reference harness, so results compare like with like.
- Held constant
- Harness, tools, data
- You receive
- Per-family diagnostics
An open benchmark for AI in finance. Alpha Bench measures how language models, agent harnesses and vertical agents perform across twelve families of investment strategy — and shows each of them how to improve, so the financial help people receive keeps getting better.
The benchmark
Twelve strategy families. One honest standard.
From value investing to tail hedging, Alpha Bench builds point-in-time research tasks, market replays, live forward tests and client conversations for each major way professionals invest.
Systems under test
One benchmark, three kinds of systems
Harnesses
For harness builders
Measure what scaffolding adds — tools, data access, memory, planning and execution — with the underlying model held constant.
- Held constant
- Model, task set
- You receive
- Component-level ablations
Vertical agents
For vertical agent teams
Evaluate a complete finance agent end to end, the way a user experiences it — including how it explains, discloses risk and says no.
- Evaluated as
- The whole system
- You receive
- Track-by-track scorecard
Why Alpha Bench
Finance is where AI has to earn trust.
Money decisions compound. A system that sounds confident but ignores costs, risk or what a filing actually says can quietly harm the people who rely on it. We think the way to make AI genuinely useful in finance is to measure it carefully, in the open — and to keep measuring as both AI and markets change.
Our approach
Point-in-time, always
Every task sees only what was knowable at the moment of decision. No look-ahead, no survivorship, no hindsight.
Outcomes and process
We score risk-adjusted results and the reasoning behind them: evidence, discipline and honesty about uncertainty.
Diagnosis, not just rankings
Every evaluation returns a breakdown by capability and failure mode, so builders know what to fix next.
A living benchmark
Tasks refresh on a rolling basis, and a live track scores decisions before outcomes exist — keeping results honest as models evolve.
Strategy families
Twelve ways professionals invest
Each family has its own tasks, metrics and failure modes. Together they cover the main schools of professional investing.
- 01Fundamental investingJudging what a business is worth, and buying it for less.
- 02Indexing & asset allocationBuilding and maintaining diversified portfolios across assets and a lifetime.
- 03Factor investing & quantitative selectionSystematic stock selection from measurable, persistent characteristics.
- 04Macro & cross-assetPositioning across rates, currencies, commodities and equities on the economic cycle.
- 05Trend & momentumFollowing prices that keep moving in the same direction.
- 06Mean reversion & statistical arbitrageBetting that stretched prices and spreads snap back.
- 07Behavior, sentiment & flowsReading what investors feel, say and do — and where the money moves.
- 08Event-drivenPricing the before and after of corporate and market events.
- 09Relative value & spread tradingTrading the price of one asset against a close relative.
- 10Carry & holding-return strategiesEarning the return an asset pays simply for being held.
- 11Options & volatilityTrading uncertainty itself — its level, shape and tails.
- 12Fixed income & creditLending to governments and companies, and getting paid for the risk.
How improvement happens
Evaluate. Diagnose. Improve. Re-evaluate.
A score on its own changes little. Alpha Bench is designed as a loop: every result points to something a builder can fix, and every fix is tested again on questions the system has never seen.
- 01
Evaluate
Systems run the same versioned tasks across four tracks: Analysis, Simulation, Live and Client.
- 02
Diagnose
Results are broken down by strategy family, capability and failure mode — from look-ahead leaks to cost blindness.
- 03
Improve
Builders receive concrete reports: where reasoning broke, which risks were missed, what a stronger answer contained.
- 04
Re-evaluate
Fresh tasks and the live track check whether improvements hold out of sample, not just on yesterday’s questions.
Latest from the Foundation
View allAnnouncementsNov 8, 2026
Introducing Alpha Bench
Why we are building an open, continuously refreshed benchmark for AI in finance — and what it will measure.
MethodologyNov 8, 2026
Why every task is point-in-time
Look-ahead bias is the most common way financial evaluations fool themselves. Here is how we design it out.
PerspectivesNov 8, 2026
From scores to diagnoses
A leaderboard tells you who is ahead. Builders need to know why — and what to change. How Alpha Bench reports results.
Help build the benchmark finance deserves.
We are looking for model labs, harness and agent builders, finance practitioners and researchers to shape Alpha Bench with us.