A Vector Set is a price channel anchored by a triad of major tops and bottoms (at least one major top and one major bottom). Think of it as a triple-decker sloped channel, a shape any Fibonacci-minded chartist will recognize:

• The core section (middle deck) spans from the anchoring bottom to the anchoring top.
• The leveled-up section (upper deck) extends above to capture resistance at higher prices.
• The leveled-down section (lower deck) extends below to capture support at lower prices.

Each Vector Set is produced by fitting a regression line through its anchoring triad. That makes the channel a formal confidence interval: the core corresponds to approximately ±0.55 standard errors of the fit, and the full leveled structure to approximately ±2 standard errors. The hand-drawn channel was always an informal confidence interval; the Vector Model makes it a formal one.

The grey lines on the chart are these individual fitted channels. VecViz fits a channel through every triad of a ticker’s major tops and bottoms, producing thousands of triad regressions per ticker, far more than any analyst could draw manually. No single line is trusted: the structure that many triads agree on is what survives. Each channel is scored for Vector Strength, which sets the weight its confidence interval receives when the channels are aggregated into the smooth support and resistance distribution described below.

Vector Strength is VecViz’s proprietary score for quantifying the support or resistance a fitted triad channel is likely to exert, and the weight that channel’s confidence interval receives when the channels are aggregated. It is calculated using three factors:

• Time proximity: channels anchored by more recent tops and bottoms have greater influence.
• Price proximity: channels terminating closer to the current price have greater influence.
• Explanatory power: channels whose fitted line better explains the ticker’s full history of tops and bottoms have greater influence. The test of quality is explanatory power: not how often price visited a level, but how much of the ticker’s structure the line accounts for.

Each triad’s ±2 standard error confidence interval enters the framework as a Gaussian component weighted by its Vector Strength. The cumulative Vector Strength between the current price and any forward price is the probability mass of that weighted mixture standing in the way. The more mass between two prices, the harder it is to get from one to the other — and the lower the probability of reaching that price, as depicted in the image below:

For more from a conceptual perspective read here and here. For practical applications read here, here, and here.

The Vector Model is VecViz’s machine-learning-based system for forecasting forward price probabilities. Its key innovation: rather than measuring price movement in dollars or percentages, the Vector Model scales price movement by the amount of support and resistance that must be traversed. A 5% move through heavy resistance is fundamentally different from a 5% move through open air.

That support and resistance is defined by a continuous distribution. Each of the thousands of triad regressions contributes one Gaussian, with its mean set by the fitted line projected to the model date and its width set by the ±2 standard error confidence band. The Vector Strength weighted aggregation of these components produces a Gaussian Mixture Model (GMM): a smooth, analytic density with a closed-form cumulative curve. Because a blend of bell curves is not itself a bell curve, the mixture can be skewed, fat-tailed, or multi-modal, so the asymmetry between support and resistance implied by the tops and bottoms is preserved rather than averaged away.

The model is trained to predict how much of this structure is likely to be traversed to the upside and downside based on a ticker’s “chart shape” profile. Chart shape includes concepts like timing of last tops and bottoms, proximity of highest tops and lowest bottoms, the angle of the fitted channels, and the distribution of the mixture’s mass above and below the current price.

For probability outputs, the mixture is anchored to the present: a single shift slides the whole distribution until its median sits at today’s price, so the resulting percentiles read as moves from where the stock actually trades. Vector Model based price probability percentiles are displayed in blue on various VecViz dashboards (whereas Sigma is displayed in red, for comparison).

Sigma is the standard deviation of returns — the foundational volatility concept used in the Black-Scholes option pricing formula and introductory finance courses. VecViz calculates Sigma using daily log returns over a two-year lookback period with a 6-month half-life decay.

The key difference: Sigma considers daily returns only. It is indifferent to their ordering and what prices they were associated, and totally unconcerned with the concept of support and resistance. VecViz displays Sigma-based probability percentiles in red alongside the Vector Model’s blue percentiles, so you can see how much the support/resistance structure affects probability estimates for a given ticker.

Yes. Earlier versions of VecViz constructed Vector Sets as Fibonacci-spaced channel lines and weighted them in part by how often price had touched those lines. The current version is very similar in substance: it relies on the same methods for identifying the major tops and bottoms that anchor every channel, and on the same downstream methods for calculating price probabilities and the V-Score.

Our motivation in shifting to the regression / Gaussian mixture model approach was improving processing speed. The efficiency gain allowed coverage to expand more than 5X as a result, from roughly 140 tickers to well over 700. Backtesting also reveals a slight performance improvement in most metrics under the current approach.

VecViz’s V-Score ranks tickers on their expected forward price return performance. We believe it to be applicable for periods ranging from 1 day to 1 year forward. The V-Score is based entirely on 13 metrics comprising Vector Model inputs and outputs and ratios thereof.

All these inputs are disclosed in the V-Score “Closest Comparables” spider chart, which shows historical ticker-dates with similar chart shapes from both top-quintile and bottom-quintile performers — helping you see what happened in analogous situations. The V-Score inputs relate to the characteristics of the fitted channels and the Tops and Bottoms that anchor them, such as the angle of the channels, and the timing of the last tops and bottoms. We sometimes refer to such metrics as being “Chart Shape” related. Vector Model probability distribution and Option Fair Value Estimate related metrics are also considered. The ensemble reads the shifted probability outputs (framed from today’s price) together with chart-shape features computed on the unshifted geometry.

The V-Score utilizes a proprietary ensemble of widely used Python machine learning model libraries to generate 6 time horizon specific scores (each ranging from -2 to +2) that sum to the “overall” V-Score (which ranges from -12 to +12). This ensemble of models is trained upon the 13 Vector Model related metrics discussed above for 30,000+ ticker-model dates spanning 100-150 tickers and 250 randomly selected model dates between June 2005 and January 2021. The only input to VecViz’s chart shape and related metrics for each ticker – model date, is a window of up to 6,500 closing prices, to the extent available.

VaR (Value at Risk) is the maximum loss you could experience at a specified future date at a given probability level. If 95% VaR is accurate, losses would exceed it only 5% of the time. VecViz shows 95D and 99D levels for both Vector Model (blue) and Sigma (red).

OaR (Opportunity at Risk) is the maximum gain you could forgo by being uninvested. It is VaR’s upside counterpart. VecViz shows 95U and 99U levels for both models.

Important: VaR and OaR estimates apply to the specified future date only — not the minimum or maximum price over the intervening period. VecViz monitors “breakage rates” (how often actual prices exceed these levels) in dedicated dashboards.

EUB (Expected Up Body) and EDB (Expected Down Body) are probability-weighted average prices within the “body” of the distribution (between the 95th percentiles up and down).

EUB tells you what price level to expect if the ticker rises but stays below the 95U tail. EDB tells you what to expect if it falls but stays above the 95D tail. Together they comprise the Vector_BodyFrcst displayed on dashboards.

These metrics attempt to capture the impact on investor returns of using the Vector Model instead of Sigma VaR and OaR metrics to size positions, assuming the investor has a fixed risk budget per ticker. They are constrained by a cap of 300% and a floor of 33.33% to ensure realistic performance parameters.

• Return on VaR Based Capital (ROVBC): Assumes position sizing is dictated by downside risk. Sigma earns the return of the ticker, and the Vector Model earns a proportionate return based on the ratio of Sigma VaR / Vector VaR. Higher ROVBC indicates the Vector Model successfully identified elevated risk, protecting capital on downside moves.
• Return on OaR Based Capital (ROOBC): Assumes position sizing is dictated by upside opportunity. Sigma earns the negative of the return of the underlying, and the Vector Model earns a proportion based on Sigma OaR / Vector OaR.
• Return on Long OaR Based Capital (ROLOBC): A variation of ROOBC oriented toward long-only, return-seeking investors. It assumes Sigma earns the return of the ticker, and the Vector Model earns a proportion based on Vector OaR / Sigma OaR.

Similar to the VaR and OaR efficiency metrics above, these capture the impact of position sizing based on the Expected Body (EUB and EDB) forecasts, utilizing the same 300% cap and 33.33% floor constraints.

• Return on Expected Down Body (ROEDB): Reflects a risk-averse strategy. Sigma is ascribed the underlying return, and the Vector Model implementation is ascribed a multiple based on the ratio of Sigma EDB / Vector Model EDB.
• Return on Expected Up Body (ROEUB): Reflects a return-focused strategy. Sigma is ascribed the underlying return, and the Vector Model is ascribed a multiple based on the ratio of Vector Model EUB / Sigma EUB.

Beyond its ticker-level offerings, VecViz provides a suite of portfolio optimization analytics that apply its support/resistance and narrative frameworks to correlation estimation and regime-based performance scoring:

• VecEvent-based correlation: Measures forward correlation between tickers based on the similarity of their VecEvent narratives — an alternative to purely price-history-derived correlation that incorporates the fundamental stories driving each ticker.
• VecViz Fingerprint-based correlation: Derives forward correlation from the similarity of tickers’ VecViz analytic “fingerprints” — their pattern of Vector Model outputs and chart-shape features — providing a price- and return-aware correlation estimate.
• VaR/OaR breakage-based regime framework: Scores expected forward ticker performance based on the historical regime of VaR and OaR breakage rates — identifying periods when a ticker’s price distribution has been systematically over- or under-estimated, and using that regime context to inform forward return expectations.

These frameworks come together in VecViz’s model portfolios. We currently offer well over 50 model portfolios, reflecting a range of ticker-level weighting caps and expected portfolio volatility caps. They are calculated using all combinations of VecViz metrics and their corresponding conventional quantitative finance counterparts (trailing 252-day, daily-return-driven Sigma ticker volatility and Pearson correlation). Each portfolio seeks to optimize expected return, estimated both on the basis of actual trailing 252-day returns and using VecViz’s metrics (V-Score, EUB, and others) in combination with our VaR breakage regime analysis.

Most traders look at a chart. Some look at the news. Few attempt to mathematically compare the two. Vector Narrative Alignment (VNA) does exactly that.

Major tops and bottoms are the natural place to make that comparison. Each turn has a story: an earnings surprise, a change in the macro regime, a shift in positioning. Because a turn can be tied to a narrative, narrative can be aligned to geometry, attaching characterized events to the very points that anchor each channel. The VNA Target Price is not a number pulled off a curve; it is the level where the narrative and the geometry agree.

VNA is VecViz’s valuation framework for determining whether a ticker is cheap or rich relative to its prevailing narrative, which we source from a leading LLM, as a baseline. Our dashboards display the VNA baseline, and an AI agent connected to our MCP server can adjust the characterization of a ticker’s narrative to reflect your own views and return a recomputed target. It compares two quantities:

• Actual Price VecLevel — where the ticker’s price currently sits within its channel, measured in standard errors of the fitted channel geometry from center. This is where we are.
• VecEvent Implied VecLevel — a score derived from LLM-characterized VecEvents, reflecting how bullish or bearish the narrative is, adjusted for whether each narrative emerged during or after the channel’s formation period. This is where the story says we should be.

VNA scores are aggregated across each ticker’s fitted triad channels on the basis of the scored support and resistance (Vector Strength) each channel represents. Because the VNA Target Price’s whole signal is the distance between price and narrative fair value, it reads the Gaussian mixture unshifted; only the probability layer is shifted to today’s price. Then, we adjust for systematic LLM bias via regression across tickers. The residual — All Ticker Adj VNA Upside (VLs) — answers: how much better is this ticker’s story than the market is giving it credit for, relative to how all other tickers are priced against their stories?

Translating this into a VNA Target Price is done by multiplying the adjusted VNA Upside by the ticker’s average Price Per VecLevel and adding it to the model date price.

A VecLevel is the price’s location within a fitted channel, measured from the channel center. The channel core is the band between the parallel lines linked to the bottom(s) and top(s) anchoring the channel.

• A VecLevel of 0.0 means the price is at the channel center.
• Values beyond ±0.5 indicate price is outside the core channel (extended or compressed).
• A VecLevel of +1.0 means the price is one core-width above center — at the center of the leveled-up channel.
• A VecLevel of −1.0 means the price is one core-width below center — at the center of the leveled-down channel.

In the fitted geometry, these readings correspond to standard errors of the triad regression: the channel core spans approximately ±0.55 standard errors, and the full leveled set approximately ±2. The standard error thereby serves as the common ruler, with the same unit measuring the VNA Target Price gap, the risk envelopes, and the chart-shape metrics.

VecLevels provide a normalized, channel-relative location metric that is comparable across tickers and time periods — the common unit of measurement that makes VNA’s cross-sectional comparison possible.

VecDates are the dates of the tops and bottoms that anchor each Vector Set — the turning points that define support and resistance channels.

VecEvents are the news themes and catalysts linked to those turning points based on timing: why did that top or bottom form? VecViz sources VecEvents using a leading LLM and characterizes each on two dimensions:

• Bias: Bullish or Bearish
• Bias Trend: Intensifying, Steady, or Waning

To our knowledge, VecViz is the only charting framework that tags narrative context to price channels — and the only framework that uses that narrative context to generate a quantified valuation.

Not all narratives influence a channel equally — when a narrative emerged relative to the channel’s formation period matters significantly.

VecViz’s theory is that a channel’s trajectory reflects the net bullish/bearish balance of the narratives active during its formation period (the span of dates between its anchoring tops and bottoms). Extrapolating the channel forward presumes that balance has been maintained.

• A bullish narrative from the formation period that is now waning reduces the net vector balance — implying some downward drift within the channel going forward.
• A bullish narrative that has intensified since formation pushes the balance more bullish — implying upward drift.
• A bullish narrative emerging after formation always contributes positively to the net balance, even if it has recently begun to wane.

These dynamics are captured in the VL_Score assigned to each Bias / Bias Trend / Timing combination:

Given VNA’s heavy reliance on recently introduced, frequently updated LLMs — and the resulting challenge of point-in-time knowledge cutoffs — rigorous back-testing is not feasible. An LLM running today cannot reliably reconstruct what it would have said about a narrative in 2019.

However, VecViz is committed to transparency. We will track and report on VNA’s forward performance from the date of introduction and regular production (May 2026) onward, published in the Reports section alongside all other VecViz performance metrics.

VecViz’s approach has several parallels to traditional quantitative finance:

Emphasis on high trading volume: VecViz places considerable emphasis on price observations associated with high trading volume by virtue of anchoring channels directly on major tops and bottoms. See the following FAQ on Tops and Bottoms for more detailed empirical analysis of this phenomenon within the framework.

Balancing Memory and Stationarity Via Sloped Channels: The framework’s emphasis on sloped channels (as opposed to flat, horizontal channels) naturally balances price and return considerations. This provides the associated benefits of capturing both market memory and stationarity. Finding a rigorous balance between memory and stationarity was a key motivation in Marcos López de Prado’s pioneering work on fractional differencing. While López de Prado’s approach is more carefully and mathematically calibrated to maximize memory while strictly preserving stationarity, VecViz’s structural use of sloped price channels functionally targets a similar conceptual middle ground between raw prices (which have memory but lack stationarity) and raw returns (which have stationarity but lack memory). This balance extends to the probability layer: the shift that anchors the probability distribution to today’s price is a minimal, location-only stationarization, removing only location while the mixture’s skew, clustering, and slope-driven asymmetry — the chart’s memory — survive untouched. The chart-shape metrics carry the same idea further: range position, top and bottom ages, and trend angles are bounded or normalized so that each means the same thing from one ticker and date to the next (the stationary property), yet each also encodes where price sits in its long history (memory), making them stationary and memory-bearing at once.

Robust Regression and Factor Model Parallels: In VecViz, the parallel between channel lines and regression lines is literal: each channel is a regression fit through its anchoring triad of tops and bottoms, and its width is the confidence band of that fit. Fitting every triad rather than privileging one line echoes robust estimators such as Theil-Sen, in which no single point can dominate the fit. Furthermore, VNA’s cross-sectional regression adjustment — regressing each ticker’s aggregate Raw VNA Upside upon its Actual Price VecLevel across 700+ tickers — is structurally analogous to a factor-model neutralization step. It removes systematic LLM bias from implied VecLevel scores, leaving the residual narrative-price gap as a clean signal. This approach is consistent with the academic tradition of Fama-French style alpha decomposition, applied here to narrative-derived valuation rather than fundamental factor exposures.

“Standard” deviation coverage: The core channel of a Vector Set — spanning from its anchoring bottom to its anchoring top — corresponds to approximately ±0.55 standard errors of the triad regression. Adding the leveled-up and leveled-down sections extends the full Vector Set to approximately ±2 standard errors of the fit, a formal counterpart to the sigma-coverage bands used in models like Black-Scholes.

Central Limit Theorem: VecViz fits thousands of triad regressions per ticker. Each can be considered an estimate of the range of price deviation based on a sampled set of turning points. Aggregating on a Vector Strength-weighted basis into a single continuous Gaussian mixture reflects the CLT’s core concept: a sufficiently large collection of samples can reveal characteristics of the full population.

Recency weighting: Ascribing greater Vector Strength to recently formed channels parallels the well-established principle that sigma-based volatility metrics are improved by applying exponential decay to the lookback window of returns. Both approaches give more weight to recent data.

A note on precision: the frameworks referenced above have exact technical meanings. VecViz does not construct information-driven bars, and it does not fractionally difference a series. The kinship is one of principle: VecViz’s choices about what to stationarize, what memory to keep, and where to read the stock at all can be understood through these frameworks, not as implementations of them.

Additional academic context:

• Academic and industry research supporting the concept of support and resistance is reviewed here.
• Academic research on neural networks and chart imaging supporting the Vector Model and V-Score is discussed here.
• The evolution toward machine learning for volatility estimation and how VecViz relates to it is discussed here.

VecViz views major price Tops and Bottoms as turning points in the net bullish/bearish vector balance driving a ticker’s price. They are the anchors of every Vector Set, the dates from which VecEvents are timed, and key inputs to the Vector Model and V-Score — several of which are direct transformations of Top and Bottom prices, dates, and the angles of the channels constructed from them.

What VecViz did not initially set out to do was incorporate trading volume. Volume is not a direct input to any VecViz model. Yet empirical analysis reveals that Tops and Bottoms tend to occur amidst elevated trading volume: across 141 tickers studied from January 2015 through January 2026, the median ticker showed a Top/Bottom Volume Ratio of 1.41× non-Top/Bottom days. Mann-Whitney testing found this ratio was significantly greater than 1.0 for 63–78% of all tickers studied, depending on confidence threshold.

This matters because trading volume has been a well-regarded quant signal for decades. Clark (1973) and Epps & Epps (1976) established that volume-weighted returns improve volatility modeling. Easley, López de Prado & O’Hara (2012) introduced the “Volume Clock” for detecting informed order flow. Subsequent research has continued to confirm volume’s predictive virtues.

VecViz’s heavy reliance on Tops and Bottoms therefore implicitly channels these virtues — without trading volume appearing anywhere in the model specification. The practical evidence bears this out:

• Tickers with statistically significant Top/Bottom Volume Ratios showed an incremental ~250bps annualized improvement in V-Score long/short performance differential versus tickers without significant Volume Ratios.
• A nearly significant improvement in 95% VaR 1-day performance relative to Sigma was also observed for high Volume Ratio tickers.

The conclusion: by anchoring everything on Tops and Bottoms — the moments when narrative forces reach a tipping point and reverse — VecViz inadvertently tapped into one of the most robust phenomena in quantitative finance. Read the full analysis here.

Convolutional Neural Networks (CNNs) excel at extracting complex, non-linear features from structured data without manual feature engineering. Gramian Angular Fields (GAFs) are a technique for transforming a price (or trading volume) time series into a square matrix amenable to CNN processing. When visualized, GAFs produce plaid-like heatmap patterns and are increasingly employed by quantitative investors for pattern recognition.

Both VecViz and Gramian Fields transform a time series into a structured representation for machine processing, but do so from different starting points. Our research shows that they capture somewhat different features. However, please note that VecViz has not yet developed a commercial product offering based on this GAF/CNN research. It remains an active area of theoretical exploration and internal development. Read the full analysis here.

The dashboards include a report that details daily, month-to-date, and year-to-date performance for most VecViz metrics, broken down by many cohort groupings (trading liquidity, sector).

We will resume production of our extensive monthly reports for each metric later this fall, after we update their disclosures. While the vast majority of the VecViz algorithm we are currently executing has been in production since 2024 and out of sample since 2022, we updated a small part of the algorithm this summer for purposes of improving processing speed. The update was purely to the algorithm: the tickers used for training, and the period they were trained upon, are unchanged.

That said, we expect those reports will cover the 700+ ticker data set going back to the start of the out-of-sample period, 1/31/2022.

A single VecViz API key entitles you to all of the following. Subscription options are at vecviz.com/signup.

• Web dashboards at vecviz.com. The charts, channels, VNA target prices, V-Scores and probability bands, rendered and updated after each market close. Built for a desktop screen.
• Our MCP server, connected directly. Any agent that accepts an API key in its configuration reads the same analytics the dashboards draw on, and can interrogate the model rather than only display it.
• Our MCP server, through Smithery. Agents that sign in only by OAuth reach VecViz through our Smithery listing.
• Our REST API. For your own code, notebooks and pipelines.
• Self-hosted implementations of the OpenBB Workspace. We expect a VecViz API key will allow the VecViz app to continue to function and receive updated analytics after each market close on a self-hosted install.

Most users combine two of these: the dashboards to see the geometry, and an AI agent connected to our MCP server to ask why a number is what it is, or what it becomes under a different set of assumptions.

For enterprise access, API integrations, or institutional inquiries, please contact us at admin@vecviz.com.

Yes. VecViz exposes its analytics through an MCP (Model Context Protocol) server. There are two ways to connect, and which one you use depends on your agent.

Direct connection. The endpoint is https://agents-vecviz.fly.dev/mcp/, authenticated with your VecViz API key passed as a bearer header. This is the shorter path, and it works with any agent that accepts a key in its configuration. As of 9/2/2026 that includes Claude Code, Cursor, VS Code, Windsurf and Gemini CLI.

Connection through Smithery. Agents that sign in only by OAuth, including Claude desktop, Claude.ai and the Claude mobile apps, cannot use the direct path because VecViz does not yet offer OAuth. These agents can reach VecViz through our Smithery listing. Smithery stores your VecViz API key and presents an OAuth-capable endpoint to your agent on your behalf.

ChatGPT is a special case. It cannot present an API key or a custom header under any configuration, so the direct path is unavailable to it, and its custom connector support requires Developer Mode, offered only on the Pro, Team, Enterprise and Edu plans. We have not yet confirmed a working ChatGPT connection through Smithery.

Use this route if your agent accepts an API key in its configuration: Claude Code, Cursor, VS Code, Windsurf and Gemini CLI all do. When you receive your VecViz API key (trial or paid), save both the server URL and the key. Then ask the agent you intend to use to:

Replace YOUR_VECVIZ_API_KEY with the key from your signup email. The agent will walk you through the configuration steps appropriate to its own environment.

To confirm the connection, ask your agent to call vecviz_catalog, which is free and returns the coverage list. Then ask for vecviz_read on SPY, which is metered. If the first call succeeds and the second fails, the connection is working and the problem is the key.

These agents sign in by OAuth only and will not accept an API key, so they connect through Smithery, which holds your key and presents an OAuth endpoint on your behalf. The steps take a few minutes because you are provisioning your own endpoint rather than using a shared one.

1. Create an account at smithery.ai. Note the namespace it assigns you, usually your username. It becomes part of your connection URL.
2. Open the VecViz listing at smithery.ai/servers/vecviz/vecviz.
3. Click Add to toolbox and choose the namespace the connection should live under.
4. Click Set up on the VecViz entry in your Toolbox and paste in your VecViz API key. The connection stays inactive until you do this.
5. Click Install and choose Claude from the client list. Smithery returns a URL of the form https://mcp.smithery.ai/your-namespace.
6. In Claude, go to Settings > Connectors, click Add custom connector, and paste in that URL.

Smithery’s install dialog also covers Codex, Cursor, VS Code, Gemini CLI, LM Studio, Cline, Zed, OpenCode, Antigravity, Cherry Studio, ChatWise, Roo Code, Amazon Q, Dust, LibreChat and Kiro, if you would rather route those through Smithery than connect them directly.

Custom connectors are not available on every Claude plan. If you do not see the option, check your subscription tier.

To view your billing history, update your payment method, or cancel your subscription, access your subscriber portal below. You’ll be asked to enter the email address you used at signup — Stripe will send you a one-time login code to verify it’s you. vecviz.com/manage-subscription.