Methodology

Both public lenses use a −3 to +3 scale, but they measure different things and the sign means something different in each.

Scores between −0.5 and +0.5 are treated as neutral throughout the site and are shown in a neutral color. A score of 0.2 is not a weak buy signal; it is the model declining to make a call.

Market Lens combines two independently scored branches: a technical read of trend, volatility and breadth, and a News & Events read built from verified, dated source documents. When both branches are usable they are weighted 60/40 respectively.

The two branches are kept visible rather than collapsed. When technical conditions are strong but news evidence is adverse, that asset class is marked contested — the disagreement is the finding, and averaging it away would destroy the most useful part of the signal.

Each piece of news evidence is a market force: a factual summary, an explicit statement of how it transmits to that asset, a direction, a time horizon, and links to the primary sources. Where a credible counterargument exists it is shown alongside, not omitted.

The single-day read is computed and presented separately from the medium-term score. They are not combined, and a divergence between them is reported rather than smoothed.

Every fair-value figure is expressed as an index where today’s market price equals 100. A median fair value of 112.5 means the model’s median estimate sits 12.5% above the current price.

The output is a distribution, not a target price. The charts show the P10–P90 and P25–P75 bands around the median precisely because the width of that range is as informative as its center. A narrow distribution around a small discount is a materially different proposition from a wide one around a large discount.

Modeled returns are compared against the current one-year Treasury yield on an investment basis. An asset with a modeled 5% return in a 4% Treasury environment is a different proposition than the same 5% against a 1% hurdle.

Where a probability is labeled base_rate_only, no regime-matched historical sample was available. In that case the site shows the model’s scenario-implied figure next to the unconditional historical base rate and labels it explicitly as uncalibrated. It is not presented as a probability of the outcome.

Portfolio Lens uses a covariance matrix estimated from five years of monthly total returns, with linear shrinkage toward the diagonal applied before optimization. Shrinkage reduces the off-diagonal terms while leaving variances unchanged, which stabilizes an estimate that is otherwise noisy in exactly the places an optimizer exploits hardest.

Simulation uses a multivariate Student-t distribution rather than a Gaussian one. Financial returns have fatter tails than a normal distribution admits, and a Gaussian assumption produces drawdown estimates that are comfortable and wrong.

Because the entire model is published as a downloadable bundle, the computation runs in your browser. Your holdings are never transmitted unless you explicitly choose to save a portfolio.

Every report is checked twice before publication: once against a schema contract, and once against a set of analytical invariants that test whether the report contradicts itself.

Those checks include, among others:

• Claimed coverage must match delivered records — a report stating 72 instruments must carry 72 instrument records.
• Percentile orderings must be monotonic in every distribution.
• Derived figures must reconcile with their inputs — a stated upside percentage must follow from the median fair-value index it summarizes.
• Every cited source must exist in the source registry, so no claim can point at a document that is not there.
• Labels must not contradict the sign of the score they describe.
• Covariance matrices must be symmetric, positive semi-definite, and consistent with their own correlation matrix and declared shrinkage.

A report failing any of these is withheld from the site entirely. Non-blocking findings are published with a visible data-quality note attached to the report rather than hidden.

These models describe conditions and relative value. They do not predict prices, and nothing here accounts for your tax position, liquidity needs, existing commitments or risk capacity beyond the parameters you supply.

Modeled returns rest on assumptions that will sometimes be wrong — about mean reversion, about the stability of historical relationships, about which risks are already priced. Correlations in particular tend to converge toward one during the exact episodes when diversification matters most, and a five-year covariance estimate will understate that.

Every report is a point-in-time snapshot with a stated data cutoff. Markets move between cutoffs. Treat the timestamp as part of the finding.