Methodology

The Verification Layer

Each data point published by Drusus carries three pieces of provenance metadata: the primary source, the time of the snapshot, and the verified prior-session close against which percentage changes are calculated. The verification layer rejects any instrument line where any of these elements is missing or stale beyond the freshness threshold defined for the relevant venue.

This is the most consequential design choice in the architecture. A briefing that omits an instrument is forgivable; a briefing that misstates one is not. The verification layer is intentionally conservative: it errs on the side of marking data unavailable rather than publishing an unverified figure.

The AI Analyst

The AI analyst in the Drusus chat interface is constructed on a large language model, constrained by a system prompt that requires it to use only the verified data context provided to it, to flag any data it is uncertain about, and to write in the analytical register of a senior professional rather than a marketing assistant.

At the Institution tier, the analyst is built on Anthropic Claude Sonnet, the highest-grade model the platform supports. At the Strategist tier and below, the analyst draws on a different model with comparable analytical capability for most questions but a more limited context window for the most complex multi-instrument analyses. The choice of model is made for each query; we do not artificially degrade the output of lower tiers.

The analyst is instructed to refuse to produce investment advice within the meaning of the regulatory perimeter, to refuse to issue buy or sell recommendations on individual securities, and to flag situations in which the user appears to be seeking such advice rather than analysis.

Scenario Modelling

The Monte Carlo simulation engine, available at the Strategist tier and above, draws on the historical return distribution of each constituent instrument from 1985 (where available) to the present, with a rolling correlation window calibrated to reflect the regime-dependence of asset correlations. The default simulation runs 10,000 paths, configurable by the user.

The engine accommodates user-specified withdrawal schedules in either nominal or real terms, with inflation modelled on the basis of the user-selected country consumer-price index. The output reports the probability of capital preservation in both nominal and real terms, the median path, and the percentiles of consequence (5th, 25th, 75th, 95th).

The principal methodological caveat is the customary one for Monte Carlo work: the historical distribution from which the simulation draws may not be representative of the regime the projection horizon will actually traverse. The output is conditional on the assumption that the past is informative about the future, and should be read with appropriate caution.

Value-at-Risk

The Value-at-Risk computation, available at the Strategist tier and above, uses the historical-simulation method with a rolling correlation window. We have deliberately not adopted the analytical (parametric) method, which assumes a Gaussian return distribution and tends to understate tail risk; nor the variance-covariance method in its simple form, which similarly underestimates the fat tails financial data exhibits.

The default lookback window is 250 trading days, configurable. The confidence level is settable, with 95% and 99% the customary defaults. Expected Shortfall (Conditional VaR) is reported alongside the headline figure for the same reason it is preferred in serious risk literature: it characterises the loss conditional on exceeding the VaR threshold, which is the figure that actually matters in the event of an exceedance.

Cross-Listing Reconciliation

For dual-listed instruments (notably the Chinese megacaps with both HKEX and NYSE listings, but also a substantial set of European and Latin American names), the platform reconciles the listings into a single instrument record with the FX-adjusted price relationship surfaced explicitly. The FX rate used is the closing spot rate of the relevant venue's session; an inter-session reconciliation is provided for users who require synchronised pricing during overlapping trading hours.

The deviation between the two listings, when normalised for FX, is itself a tracked analytical object. Persistent deviation outside the historical range is flagged for user attention, since such deviations frequently reflect either a settlement-related friction or a divergent regulatory consideration in one of the two jurisdictions.

Drusus Daily Editorial Pipeline

Each edition of Drusus Daily is produced through a four-stage pipeline. First, a data snapshot is assembled from the platform's sources, each figure carrying its provenance metadata. Second, the verification layer rejects any incomplete or stale data lines. Third, the verified context is passed to a language model constrained by a system prompt that requires use of only the provided data and instructs the editorial voice. Fourth, an automated review pass examines the generated prose for any reference to a figure not present in the verified context, rejecting the edition if such reference is detected.

Editorial oversight is provided by Gravenos. Where the automated review surfaces an issue that requires human resolution, the edition is held until the issue is addressed. The platform will publish a partial edition with a transparent note rather than a complete edition containing unverified content.

Citation Guidance

Where the platform is used as a source of data or analytical context in academic or professional research, we ask that it be cited as follows: Gravenos, Drusus Platform, gravenos.com, with the date of data retrieval recorded. Where the methodology contributes substantively, please cite this page directly. Where Drusus Daily content is used in research, please cite the relevant edition by date.