Vero logoVERO

Case study

Mini-FRED: Deterministic finance Q&A benchmark

Offline evaluation with database-grounded truth. Shows how reliability improves across versions.

This is a benchmark scorecard, not a live system claim.
Vero's rigor of using MVES (Minimum Viable Evaluation Suite) is a deterministic, end-to-end evaluation harness for data agents. It turns natural-language questions into structured expectations, runs your agent against a local, versioned data warehouse, and scores answers with transparent, reproducible checks. No hidden APIs, no flaky metrics-just clear pass/fail signals, failure breakdowns, and reports you can track over time.
Mini-FRED is a 5-year collection of FRED data used to answer natural-language finance questions with deterministic transforms: point (raw value at a date), YoY (year-over-year % change), MoM (month-over-month % change), MA (moving average over N periods), max (maximum in a window), and min (minimum in a window).

Executive summary

  • Reliability improved from 65.9% (v1) to 78.2% (v5), a +12.3 point lift.
  • Primary remaining issue: Transform confusion (intent parsing drives the wrong computation path).
  • Next improvements focus on rules-first transforms, a stricter output contract, and clarifying questions for ambiguous phrasing.
ProgressionPass rate
v1
65.9%
v2
69.8%
v3
69.8%
v4
66.8%
v5
78.2%
View the full repo →

Offline MVES benchmark • 560 questions

Mini-FRED Agent v1

Baseline deterministic agent scored against DuckDB-grounded truth.

Benchmark Index (proxy)

657.4

Success Rate

65.9%

Critical flags (assertion)

441

Pass / Fail

369 passed

191 failed

Top failure modes

Wrong computation path · 181Transform confusion · 142Output numeric formatting · 60

Baseline parser + deterministic DuckDB truth checks.

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

1 / 6Agent version

Mini-FRED Agent v1

Baseline deterministic agent scored against DuckDB-grounded truth.

Executive takeaway: Baseline reliability with clear transform errors.

What changed

  • >Baseline parser + deterministic DuckDB truth checks
  • >Single-pass transform selection
  • >No intent normalization yet

Primary remaining issue

Wrong computation path (misidentifies the requested transform, so correct data but wrong calculation).

Offline MVES benchmark • 560 questions

Mini-FRED Agent v2

Improved parsing to reduce date handling and extraction errors.

Benchmark Index (proxy)

696.8

Success Rate

69.8%

Critical flags (assertion)

339

Pass / Fail

391 passed

169 failed

Top failure modes

Wrong computation path · 159Transform confusion · 120Output numeric formatting · 60

Improved date parsing + stricter value extraction.

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

2 / 6Agent version

Mini-FRED Agent v2

Improved parsing to reduce date handling and extraction errors.

Executive takeaway: Higher success rate; computation path still dominant.

What changed

  • >Improved date parsing + stricter value extraction
  • >More explicit transform coercion
  • >Cleaner numeric extraction fallback

Primary remaining issue

Transform confusion remains (better parsing, but still picks YoY vs MoM incorrectly in noisy phrasing).

Offline MVES benchmark • 560 questions

Mini-FRED Agent v3

Retrieval + parsing refinements to stabilize answer formatting.

Benchmark Index (proxy)

696.8

Success Rate

69.8%

Critical flags (assertion)

339

Pass / Fail

391 passed

169 failed

Top failure modes

Wrong computation path · 159Transform confusion · 120Output numeric formatting · 60

Retrieval + parsing refinements; stabilized outputs.

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

3 / 6Agent version

Mini-FRED Agent v3

Retrieval + parsing refinements to stabilize answer formatting.

Executive takeaway: Stability improved, but core failure types persist.

What changed

  • >Retrieval + parsing refinements; stabilized outputs
  • >Reduced ambiguous transform collisions
  • >More consistent series selection

Primary remaining issue

Wrong computation path persists (formatting stabilized, but intent errors still dominate).

Offline MVES benchmark • 560 questions

Mini-FRED Agent v4

Guardrails added for windows, dates, and refusal correctness.

Benchmark Index (proxy)

666.6

Success Rate

66.8%

Critical flags (assertion)

394

Pass / Fail

374 passed

186 failed

Top failure modes

Wrong computation path · 176Transform confusion · 120Output numeric formatting · 78

Window/date guardrails; more refusal correctness.

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

4 / 6Agent version

Mini-FRED Agent v4

Guardrails added for windows, dates, and refusal correctness.

Executive takeaway: Guardrails helped, but wrong paths still frequent.

What changed

  • >Window/date guardrails; more refusal correctness
  • >Refusal criteria made explicit
  • >Tighter date-window matching

Primary remaining issue

Edge-case date parsing (improved windows/Moving Average, but some ambiguous date prompts still fail).

Offline MVES benchmark • 560 questions

Mini-FRED Agent v5

Local Phi-4 intent normalization reduces transform ambiguity.

Benchmark Index (proxy)

780.9

Success Rate

78.2%

Critical flags (assertion)

301

Pass / Fail

438 passed

122 failed

Top failure modes

Transform confusion · 128Wrong computation path · 119Output numeric formatting · 54

Local Phi-4 intent normalization for transform detection.

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

5 / 6Agent version

Mini-FRED Agent v5

Local Phi-4 intent normalization reduces transform ambiguity.

Executive takeaway: Best reliability so far; transform confusion remains.

What changed

  • >Added local Phi-4 intent normalization for transform detection
  • >Reduced transform ambiguity on edge phrasing
  • >More consistent transform labels downstream

Primary remaining issue

LLM still misses some nuance (Phi-4 helps, but certain change vs level cues still misfire).

Planned improvements roadmap

What's next

Potential upgrades focused on transform clarity and output contracts.

Benchmark Index (proxy)

—

Success Rate

—

Critical flags (assertion)

—

Pass / Fail

—

Focus areas

Transform clarity & output contracts

BenchmarkIndex = round((passRate * 10) - (criticalFailures/totalCases) * 2, 1)

6 / 6Roadmap

What's next

Potential upgrades focused on transform clarity and output contracts.

What changed

  • >Rules-first transform lexicon (map phrases like "annual swing" -> YoY, "month-to-month" -> MoM)
  • >Output contract: enforce {series_id, transform, date/window, value} and single numeric
  • >Confidence gate: ask a clarifying question when date/window/transform is ambiguous
  • >Add more transforms step-by-step later (avg, median, CAGR, z-score)
  • >Expand eval coverage for ambiguous phrasing + adversarial paraphrases

Focus

Transform ambiguity + output contract discipline

Methodology

How the benchmark is computed

Benchmark Index is an internal proxy score, not ROI.

DuckDB computes truth from cached mini-fred series; the agent parses intent and selects a transform; evaluation checks the reported value within tolerance.