DELPHI: A Daily Longitudinal Survey of Machine Beliefs About a Real Asset Class

1. Introduction

Expectation surveys occupy a central place in macro-finance because beliefs move markets independently of fundamentals. The ZEW Indicator of Economic Sentiment has polled human financial analysts monthly since 1991; central banks run professional-forecaster surveys precisely because the distribution of expectations — not only its mean — carries information.

A new class of market participant has appeared. Large language models draft investment memos, screen markets, and answer the question “should I buy property in Spain?” millions of times a year. Their beliefs propagate into human decisions through every such interaction. Three properties make these beliefs worth recording systematically: they are influential (model-mediated research is a growing share of investment workflow), they are heterogeneous (different model families produce materially different quantitative beliefs, as our first panels demonstrate), and they are perishable — a model's belief on date t is only observable on date t. Unlike price data, the series cannot be backfilled, which makes a continuous record valuable in proportion to its length.

DELPHI — named for the Delphi survey method, whose round-one structure it implements with machine panelists — is, to our knowledge, the first instrument to record these beliefs daily against a fixed question bank with pre-registered resolution criteria.

2. Methodology

Question bank. Twelve forward-looking quantitative questions about European residential property, each typed as a probability (0–100%), a percentage change (−10%…+10%), or a 0–100 scale rating; each tagged with a directional sign (whether a high answer is bullish or bearish for the asset class), a horizon in months, and a resolution source — a named public statistic (ECB MFI interest-rate statistics; Eurostat house-price index; national statistics offices) against which the question resolves at horizon. The bank is version-controlled; any change increments the published version.

Panel and elicitation.The launch panel comprises three models from two independent providers, intentionally mixing retrieval-augmented and parametric-knowledge panelists. Each question is posed in a fresh context with an answer-only instruction (a single number, no reasoning) to suppress format drift. Panelists never see one another's answers — a true Delphi round one. The operator's own analytics never participate: the referee does not play on the scoreboard.

Aggregation. Per question: consensus = median; dispersion = max − min. Per day: the Consensus Index is the mean of bullishness-normalized answers (50 = neutral, higher = collectively bullish for European property); the Disagreement Index is the mean dispersion. Medians and ranges are preferred for robustness with small panels.

Integrity. Every run is event-sourced and replayable. Daily artifacts are committed in a Merkle root, timestamped under RFC 3161, and anchored to a Zenodo DOI. The full per-model, per-question, per-day record is public via API and mirrored daily to a public git repository whose commit history independently witnesses the series.

Resolution and calibration.At each question's horizon the realized outcome is read from the pre-specified source. Probability questions score by Brier score; quantitative questions by absolute error. Accumulating resolutions yield per-model calibration curves — a public track record of machine judgment, complementing knowledge benchmarks with a measure of foresight.

3. First-panel findings

The inaugural panel (2026-06-10) opened at Consensus Index 53.3 (mildly bullish) with Disagreement Index 19.9. The widest split concerned the probability of ECB rate cuts within six months: 25% versus 72% — a 47-point spread between frontier models on the single most consequential variable for the asset class. Persistent, attributable inter-model disagreement of this size on a well-posed question is itself a finding about the epistemic state of deployed AI systems. We further observe round-number anchoring in a smaller panelist (identical values returned across unrelated questions) and systematic differences between retrieval-augmented and parametric beliefs. The live series at avenaterminal.com/delphi supersedes this section daily.

4. Limitations

Panel size is small at launch; the architecture admits any model exposing an API. Answer-only elicitation trades reasoning transparency for comparability; alternative elicitations are a planned ablation. What an LLM “believes” is operationalized strictly as its answer under the fixed protocol — the protocol-conditional belief, which is precisely the quantity that propagates to users. Providers update models; version strings are recorded per response, making transitions visible breaks rather than silent drift.

The central property bears repeating: the series cannot be reconstructed retroactively. Whatever its eventual scientific use — machine herding, calibration, the transmission of model beliefs into prices — the prerequisite is that someone recorded the beliefs at the time. That is what DELPHI does, daily.

Data availability

Live instrument: /delphi · JSON: /api/v1/delphi · RSS: /feed/delphi.xml · Daily git mirror: github.com/HenrikKolstad/avena-data · DCAT-AP: /catalog.jsonld · License CC BY 4.0 · DOI 10.5281/zenodo.19520064. Companion benchmark (PLAB): /benchmark.