QuantDesk

An Isolation Forest reads six behavioural features per trading day — return, relative volume, absolute return, Money Flow Index, an on-balance-volume z-score, and intraday range. Days that don't look like other days get flagged, labelled Accumulation or Distribution by a four-way vote of money-flow indicators, and scored 0–100.

Alongside it runs a CUSUM change detector, which exists to catch what the first model structurally cannot. An institution building a position splits its order across weeks precisely so that no single day stands out — a per-day outlier detector can only ever catch the impatient buyer. CUSUM accumulates small deviations instead, so a long run of unremarkable days trips a threshold none of them would alone. On AAPL it surfaced a 27-day accumulation running at 1.04× average volume: invisible to any volume-spike rule.

Five sections behind a horizon selector, ordered longest-first so reading left to right walks from the strongest evidence to the weakest. The long horizon leads with a checklist — 200-day average, Faber's 10-month rule, 12-1 momentum, ADX, Hurst exponent, trend-line fit, 52-week position, drawdown survivability — each line stating which way it points and why.

Then the table that actually matters: every overlapping holding period in the history, so "worst 3-year window" replaces a headline CAGR that only describes one lucky start date. Then what holding it cost — maximum drawdown with depth, duration and recovery, plus the Ulcer index, which correctly scores a long shallow grind as worse than a sharp fall, because that is how it feels to hold.

Setups are pre-registered and checked in order, and "none of them is present" is the most common answer. A test fires the detector across twenty random walks and fails the build if a majority produce a trade. Candlestick patterns are detected, graded weak, and firewalled — none is permitted to place an entry, stop or target.

Three models routed automatically by sector: discounted cash flow for most companies, a dividend discount model for banks and insurers, and residual income for financials with no usable dividend data. Each runs a five-year projection plus terminal value, then a 10,000-draw Monte Carlo to produce a range rather than a single number.

The panel also solves the model backwards: what growth rate would make today's price correct? On AAPL that came out at 37% a year for five years against a 10% assumption — a claim about the world you can agree or disagree with, rather than a "fair value" a reader has no basis to judge. It is stated as conditional on the other inputs, because across plausible discount rates the same price implies anywhere from 24% to 42%.

Three published accounting screens computed from filings the app has already fetched: the Piotroski F-Score for fundamental trend, the Altman Z''-score for distance from distress, and the Beneish M-Score for accruals resembling companies later found to have manipulated earnings. The Altman variant is the emerging-market one, so an IDX listing and a US one land on the same scale.

The breadth half of the workflow batch-downloads up to 250 symbols in a handful of upstream calls and scores each on seven price-and-volume signals. Every signal becomes a cross-sectional percentile before anything is combined, because "top decile of this scan" is a claim the data supports and "82/100" is not.

The panel then does what composite scores usually hide: it reports the measured rank correlation between every pair of signals and the participation ratio of that matrix — how many genuinely independent signals the composite is really averaging. On a real Dow scan, momentum and trend correlate at +0.98 and seven columns carry about 3.2 signals' worth of independent information. That number is printed in the panel header rather than buried.

In the largest type on the page, on every run, the confluence rail asserted that the four lenses rest on two independent bodies of data — and therefore that when the two agree, the agreement is not one fact counted twice. Nothing had ever measured it. Both the rail and the explainer admitted as much, and gave the same excuse: the ranking panel can measure its own overlap because a scan gives it a cross-section, and a single ticker does not. That is true of a request and false of a script.

The statistic is Cohen's kappa — observed agreement minus the agreement each lens's own habits already supply. Raw agreement is uninterpretable here for the same reason a raw screener hit count is: a lens calling 70% of companies cheap and one calling 70% sound land on the same label 58% of the time while sharing nothing at all. Kendall's tau-b runs alongside it, because kappa asks whether two lenses reach the same label and tau-b whether they order the same way. Intervals are bootstrapped over names rather than taken from a closed form that conditions on marginals which are themselves estimates.

It ran on all 168 deduplicated names of the four index universes, each pushed through the production payload builders — the same API calls the ticker bar actually sends, so the number describes this app rather than a lookalike. A lens that could not read does not vote: a bank's refused accounting screens recorded as neutral would manufacture agreement with every other lens that happened to be quiet.

The claim survived. The panel now reports it in place: measured across 121 names in the Dow Jones Industrial Average and the Nasdaq-100, the price family and the filings family agree no more often than chance would put them there, κ = +0.05. Across all four index universes — 168 deduplicated names — the figure is +0.03. So agreement between them really is two facts rather than one counted twice.

What ships is the whole table, not the headline: every pair of lenses with its raw agreement rate, the rate chance alone would produce, κ, a bootstrapped 95% interval, τb and the number of names behind it. Three of the six pairs come out slightly negative. It is stamped with the date it was measured and names the script to re-run, because a rate like this decays with the lists it was taken on.

Flow and Trend — the pair deliberately collapsed into a single vote because they read the same price-and-volume series — came out at κ = +0.07, on an interval running from −0.02 to +0.17 that straddles zero. From the other end, the participation ratio says the four lenses carry 3.72 lenses' worth of independent information, not the two they are counted as.

The grouping was left alone anyway, and that is the part worth keeping. A kappa near zero cannot distinguish a reading that carries separate information from one that is mostly noise — both are uncorrelated with everything. The Flow lens's own event study returns no significant effect on most tickers. A lens that is independent because it is uninformative has not earned a vote of its own, so the conservative collapse stands and the panel explains why rather than quietly claiming credit for the extra independence.

Above the tabs, a plain-English summary reports what the lenses agree on, names where they disagree, states what it cannot tell you about this particular company, and lists what to check next. It is deliberately prose and not a number: collapsing four disagreeing models into one score discards every finding the rest of the app works to establish.

Each number carries an info affordance that says what it measures in plain English, whether this value is good or bad and why, and what would make you act differently — or admits that nothing would. A Guided/Full toggle defaults to Guided, folding expert tuning controls behind one labelled disclosure so the interface is legible to someone who has never read a DCF.

A long-term investor shown "Stochastic 82, overbought" next to "price is above its 200-day average" has been handed two statements of very different weight, presented identically. Grouping by horizon is a small change that removes a real category error.