How VecViz Expanded
Coverage > 5×

01

How It Works

Start with the map.

Chart structure and characterized narrative run in parallel and converge on the V-Score (timing), the price probability (risk), and the VNA Target Price (valuation). Narrative reaches valuation directly, and risk through a VecEvent-based correlation that aggregates single-name probabilities into portfolio risk.

From a stock's major tops and bottoms, and the channel structure they anchor, two things follow. The first is a set of long-term chart geometry features: where price sits in its historical range, how long since the last major top and bottom, how steep and how one-sided the strongest trends are. The second is an estimate of cumulative support or resistance: how much of the channel structure lies between the model-date price and any given price above or below it.

We use the first in machine-learning models that attempt to forecast price movement, scaled in units of the second. That process generates our price probability estimates. A similar model, which also takes the price probability estimates as inputs, generates the V-Score. On the narrative side, an analyst working with an agent characterizes the events that moved the stock as VecEvents; aligning those to the channel geometry produces the VNA Target Price, and a VecEvent-based correlation lets the single-name probabilities aggregate into portfolio risk.

02

What Changed: The Channel Specification

Coverage went from 140 tickers to more than 700. One change made that possible, and very little else moved.

The legacy method drew Fibonacci channels through selected tops and bottoms, at most about 325 per name. The current method fits a regression, with its confidence interval, through every triad of tops and bottoms, roughly 2,028 per name, and blends the results into one Gaussian mixture. It is about 5× faster to work with the 2,028 regressions than with the 325 Fibonacci channels, and that speed is the whole story of the coverage expansion.

Almost everything else is the same, including the training data behind the machine-learning models. The only other changes are two. We no longer consider horizontal channels anchored by a single top and a single bottom; every channel is now a triad. And we shift the resulting distribution so that it is centered on the current price, for the probability and timing outputs. The rest of this note walks the method in that light: what stayed, what changed, and why the change is modest in substance even as it is large in reach.

03

Signal, Not Noise

None of this changed. It is worth restating, because it is the foundation.

The Vector Model starts from a stock's major tops and bottoms, the high-volume turning points where conviction visibly changed hands. These are the highest signal-to-noise events a chart offers: a major top or bottom is where a prevailing story met enough disagreement to reverse. Building on them, rather than on daily wiggles, means the model is reading structure instead of chasing noise.

This instinct has a well-known cousin in market microstructure. A calendar clock treats every day as equally informative, oversampling the quiet stretches; the “volume clock” of Easley, López de Prado, and O'Hara1 argues for sampling by information arrival instead. Reading a stock at its major turns is the long-horizon version of that idea. We do not select turns by volume, yet across our coverage we have found that VecViz tops and bottoms tend to fall on markedly higher-than-usual volume, a median near 1.4 times an ordinary day, corroboration that the turns land where activity concentrated.

Just as important, major turns are narratable. Each one has a story, and because a turn can be tied to a narrative, we can align narrative to geometry, attaching characterized events, sourced by analyst and agent together, to the very points that define the channel. This is the basis of Vector Narrative Alignment, and it is the same today as it was before the change.

04

Every Triad, Not One Line

Speed changed how many lines we could afford to draw, and that changed which lines we drew.

In the Fibonacci era two things limited us. The algorithm was slow enough that the number of channels per name had to be rationed. And trading-room convention shaped which top-and-bottom structures a channel would be drawn from at all: the pairs and triads a chartist would recognize, not the full set of combinations the turns actually admit.

The regression-based algorithm is fast enough to take the opposite view. Consider every combination, two tops and a bottom or two bottoms and a top, in any order, and let the scoring of the support and resistance each channel represents impose the selectivity. A channel whose line explains the stock's history of turns carries weight; one that does not is outvoted. That is the Theil-Sen instinct2, the median of every pairwise slope rather than any one favored line, applied to triads so that each fit carries a width as well as a slope.

Every triad casts a vote. A line is fit through each triad of major tops and bottoms, from its earliest pivot forward. Carried to the model date, the ensemble piles into a distribution of where price could be today. No single line is trusted; the structure that many triads agree on is what survives.

05

Two Channels, Three Pivots

The same three pivots, two different rulers. Here is exactly how they differ.

Take any triad: a matching pair, two tops or two bottoms, and a lone pivot of the other kind. A sloped Fibonacci channel and a regression with its confidence interval can both be drawn through the same three points. VecViz's Vector Set was the former; the mixture is built from the latter.

Slope. The Fibonacci baseline runs through the matching pair, so the pair alone fixes the slope; the lone pivot sets only the width. The regression lets all three points vote. The two slopes coincide only when the lone pivot sits exactly midway between the pair, and can never coincide when it sits outside the pair's span. Everywhere else the channels fan apart, and the gap grows with every session projected.

Location. The two constructions share exactly one point. Because the regression passes through the centroid of the three pivots, it crosses one third of the way from the pair's rail toward the lone pivot, at the mean of the three dates. The statistical channel hangs a third of the way up the Fibonacci body, not halfway.

Width. The Fibonacci width is the vertical distance from the pair's rail to the lone pivot. The regression's width is the standard error of the fit, on the triad's single degree of freedom. The two are tied by a ratio that depends only on where the lone pivot sits along the pair's span: one standard error is between 0.707 and 0.816 of the Fibonacci width whenever the lone pivot lies between the pair, so the ±1 SE band spans 1.41 to 1.63 widths, wider than the Fibonacci body and hung off center. When the lone pivot lies outside the pair, the ratio falls below 0.707 and the Fibonacci width becomes an extrapolation from a rail the lone pivot never touched.

Levels. The Fibonacci ladder re-adds one width at fixed ratios. The regression ladder re-adds one standard error at chosen multiples. Read on the Fibonacci scale, the regression's rungs land at ratios that are not Fibonacci numbers, and because the slopes differ, they tilt.

Two channels from the same three pivots, shown here as two lows and a high. The Fibonacci channel (slate) pins its slope to the matching pair; the regression (orange) lets all three vote and bands by standard error. The ring marks the one point the two always share.

Close cousins, then, but not the same object, and the regression is the disciplined one: it uses all three points, its width is a fitted error rather than a drawn distance, and its unit is the standard error. That is why everything downstream is measured in standard errors, and why the switch cost nothing in substance.

06

Chart Shape, in Standardized Terms

These are computed exactly as before.

The standardized chart-shape metrics are the geometry features the machine-learning models consume: where price sits relative to the highest top and the lowest bottom the stock has made, how long since each, how steep and how one-sided the strongest trends are, each placed on a common scale so it means the same thing from one stock to the next. They are derived from the tops and bottoms and the channel structure in the same manner as before, on the unshifted geometry. The models that read them were not retrained on new inputs, because the inputs did not change.

A few of the chart-shape metrics on their shared standardized scale, unchanged by the switch in channel specification.

07

Smoother Distributions, Anchored to Today

Each triad projects a trajectory with a band. Summed, they are a distribution. Slid onto today's price, they are the odds.

Each triad is a small Gaussian forecast: a projected level from its line, and a width from its confidence band. We weight each by how well its line explains the stock's full history of turning points, and we sum them. Because the components sit at different levels with different widths, their blend is not a single bell curve. It can be skewed, shoulder-heavy, or fat-tailed, so the asymmetry between support and resistance that the turning points imply is preserved rather than averaged away. Where the Fibonacci method produced a binned histogram of projected levels, the mixture is smooth and continuous, with a closed-form density.

A real one. AMAT, model date 2026-06-29: the pre-shift mixture (blue) is multi-bumped and right-skewed, the weighted sum of thousands of triad Gaussians. The orange curve is the same shape translated, which is the anchoring step below.

The distribution of trajectories has a shape, but no home relative to today. This is the second of the two smaller changes: we solve for a single shift that slides the whole distribution until its median sits on today's price. For the odds of where price goes next, the question should be framed from where the stock actually trades, and once the median is pinned there, the only thing left to settle is migration, how far price is apt to travel and in which direction. That is exactly what the standardized metrics answer, in units of cumulative support and resistance, and it is why the models could keep their training data: the inputs and the scale are the same, only the distribution's location was fixed.

There is a useful way to read this restraint. To make a distribution comparable across stocks and dates it has to be stationarized, and the blunt way, differencing price into returns, purges almost everything a chart remembers. We take the opposite care, in the spirit of López de Prado's argument that stationarity and memory should be balanced rather than traded away.3 The shift removes only location. The skew, the clustering, the slope-driven asymmetry, and the chart-shape metrics are memory, and we keep them.

A small shift. NFLX: pre-shift median 77.98 against a price of 73.78, a slide of only −4.20; the two curves nearly coincide.

A moderate shift. CVS: pre-shift median 85.81 against a price of 103.58, a shift of +17.77, the same distribution slid over. AMAT, above, is the extreme case at +482.43. The pre-shift median is the structural geometry the target price measures against, in standard errors and given the narrative, not a price target.

08

Three Answers: Timing, Risk, Valuation

From this single reading the model produces one answer for each question an investor actually asks:

• Timing: the V-Score, a forward ranking by horizon from a model that reads the standardized chart-shape metrics together with the price probability estimates.
• Risk: the Price Probability Forecast, an asymmetric, time-horizon-specific, support-and-resistance-aware distribution read from today's price, with the VaR, OaR, EUB, and EDB envelopes; at the portfolio level, the single-name distributions aggregate through a VecEvent-based correlation measure.
• Valuation: the VNA Target Price, which reads the unshifted trajectories and reports where price sits against its narrative-aligned fair value, in standard errors.

The three contextualize one another: a valuation gap means more when the timing rank is strong and the risk envelope leans the same way. All three are the same outputs they were before the change.

09

In Short

More than 5× the coverage, from one change.

The tops and bottoms are the same. The chart-geometry features are computed the same way, in the same support-and-resistance units, and the machine-learning models that turn them into price probabilities and the V-Score were trained on the same data. What changed is the channel specification: a mixture of roughly 2,028 regression confidence intervals in place of at most 325 Fibonacci channels, about 5× faster, and because it can afford to consider every triad, it lets the scoring do the selecting. Two smaller changes came with it, no more horizontal single-top, single-bottom channels, and a distribution centered on today's price for the probability and timing outputs. Same three answers, timing, risk, and valuation; a sharper ruler underneath.