AI stock ratings are calculated by collecting inputs, transforming them into comparable measures, combining those measures under a model or formula, and mapping the result to a score or label. There is no universal calculation. A rating may summarize a generated investment thesis, compare financial factors with sector peers, or express a model's estimated likelihood of beating a benchmark over a stated period.
My first check is to identify which of those jobs the rating performs. The word “Buy” can appear beside very different calculations. You need the target, horizon, inputs and mapping rule before you can compare two labels meaningfully.
For this guide, I traced 766 archived iPulse AI equity consensus records back to 5,362 unique contributing rating records. The model-assigned labels were populated, but the selected per-report formula-label and subjective-confidence fields were empty. That finding matters because a missing field cannot be treated as a second opinion, a zero confidence reading or evidence that two calculations agree.
I am the founder of iPulse AI, the Open Agentic Investment Research Platform. This is my practical inspection framework, supported by a bounded historical extraction and official provider documentation. It is research education, not an independent performance comparison or a recommendation to trade a named stock.
Start with the target the rating is trying to measure
Before inspecting weights, write down the question the system answers. Is it comparing companies' financial characteristics? Estimating benchmark-relative performance? Describing a plausible absolute return? Summarizing an investment argument? These objectives require different evidence and evaluation rules.
For example, a company can have improving financial characteristics and still face an unattractive valuation. It can generate a positive absolute return while trailing its benchmark. A report can describe an appealing five-year scenario without supporting a strong near-term rating. Those statements can coexist because they answer different questions.
The horizon is part of the target. A three-month benchmark-relative output should not be compared directly with a five-year business scenario. Converting both to a five-star display changes their appearance, not their meaning. Preserve the provider's original units and labels before making any comparison.
It also helps to identify the covered universe. A sector-relative assessment evaluates a company against a particular peer group. A ranking within a screened population answers another question. A missing company may be outside coverage rather than poorly rated. Do not silently turn unavailable coverage into a negative signal.
Three different paths from inputs to a rating
| Rating method | What enters the calculation | What the output represents |
|---|---|---|
| Model-generated judgment | Evidence and context supplied to a research model, plus its task and instructions | A generated interpretation or categorical view requiring claim-level inspection |
| Quantitative factor assessment | Defined financial, market and estimate measures, often compared with peers | A score under a specified factor and aggregation method |
| Predictive benchmark model | Features and historical learning data tied to a defined outcome window | A provider's estimated benchmark-relative probability or score |
These are categories for inspection, not a league table. A platform can combine them. The important distinction is which mechanism produces the particular number you are reading, rather than whether the product uses AI somewhere in its workflow.
Seeking Alpha's official Quant Ratings guide describes sector-relative comparisons across Value, Growth, Profitability, Momentum and EPS Revisions. It assigns factor grades and an overall rating on a 1.0–5.0 scale. That is a quantitative assessment; it should be distinguished from an analyst's written thesis or a generative report.
Danelfin's calculation guide describes combining signals in decision trees to estimate the probability of beating the market over the next three months. It explicitly says the overall AI Score is not an average of the Technical, Fundamental and Sentiment subscores. These are the provider's method descriptions, not probabilities or investment results I independently validated.
When comparing these systems, my next question is what evidence would test their stated objective. A factor assessment needs its definitions and comparison population. A probability estimate needs calibration evidence over the stated window. A generated judgment needs its supporting facts, assumptions and opposing case. One kind of test cannot stand in for all three.
Follow the calculation through inputs, transformations and weights
Inputs can include prices, reported financial figures, analyst estimates, historical volatility, forecast paths and qualitative evidence. Record the source period and availability date. A number published after the research cutoff should not be treated as information the original system already possessed.
Transformation makes inputs usable together. A ratio requires numerator and denominator definitions. A growth measure requires a period. A return needs a starting price, endpoint and treatment of dividends. Standardizing variables or comparing them with peers changes the meaning of the input; it does not merely tidy the data.
Weights determine how the transformed inputs contribute. Some are explicit coefficients in a formula. Others arise from a trained model's structure and interactions. A list of important features is not necessarily a complete recipe for rebuilding the score. Ask which parts are disclosed, which are learned, and which cannot be reconstructed from the available documentation.
Avoid assuming that equal-looking subscores receive equal weight. One measure may be a diagnostic shown for explanation, while another enters the actual scoring equation. A displayed risk or agreement measure can be useful without being an independent vote. The provider should explain that relationship.
For a formula, verify units before doing arithmetic. Percentages represented as 12 and 0.12 require different handling. A forecast's cumulative return and annualized return are also different inputs. Our CAGR and total-return guide explains why averaging and compounding cannot be interchanged casually.
A worked calculation with hypothetical inputs
The archived iPulse AI configuration records a 3.6% cash hurdle, a 35% dispersion discount capped at 60% of excess-return magnitude, and a 35%/65% blend of volatility-aware and return-anchor terms. I checked these parameters in the selected records rather than assuming a current documentation page defines every historical edition.
To make the arithmetic inspectable, suppose two hypothetical grouped annual return scenarios are 6% and 18%. They are invented teaching inputs, not forecasts for any company. Their mean is 12%, and their mean absolute deviation from that mean is six percentage points.
| Calculation step | Hypothetical arithmetic | Result |
|---|---|---|
| Mean annual return | (6 + 18) / 2 | 12% |
| Excess above cash hurdle | 12 − 3.6 | 8.4 percentage points |
| Dispersion discount | Smaller of 0.35 × 6 and 0.60 × 8.4 | 2.1 percentage points |
| Adjusted excess return | 8.4 − 2.1 | 6.3 percentage points |
| Blended direction term | 0.35 × (6.3 / 20) + 0.65 × (6.3 / 15) | 0.38325 |
For the last row, I assumed 20% historical volatility, a 15% return anchor and a high-volatility multiplier of one. Those assumptions are part of the example. Applying the recorded bounding transformation, tanh of half the direction term, gives approximately 0.1893.
That is an intermediate direction term, not the final rating. Agreement, forecast-path instability, event-risk adjustments, rounding and economic guardrails can affect later stages. I have deliberately stopped the example before those stages because their required inputs have not been specified. Multiplying the intermediate value by 1,000 and presenting it as a complete score would omit part of the calculation.
The useful lesson is the sequence: a raw return scenario becomes an excess-return measure, dispersion changes its strength, normalization changes its scale, and a bounding function changes sensitivity. The calculation is not equivalent to counting favorable labels. A more detailed technical reference is the iPulse AI Consensus Score methodology.
What I found in 5,362 archived rating records
I selected every original Batch 7 equity consensus record in the non-adjusted voice mode, covering 383 companies at one-year and five-year summary horizons. That produced 766 consensus records. Each company's two summaries referenced the same 14 unique contributing tasks, so I counted each contributor once across horizons.
All 5,362 task IDs matched an active archived rating record with the same stable asset identity. I selected the latest active metadata version per task and checked unique IDs, joins, category values and missing fields. These are historical model outputs: underlying price anchors are dated 18 September 2026, contributor snapshots span 20–21 September, and consensus scoring ran on 21 September. The extraction date is 12 October, not a new forecast date.
| Stored model-assigned label | Unique contributor records |
|---|---|
| Strong buy | 744 |
| Buy | 2,243 |
| Neutral | 1,187 |
| Partially sell | 854 |
| Sell all | 334 |
| Total | 5,362 |
These are category counts, not measured success rates or recommendations. All selected rating records carried a five-year report horizon label. That does not turn the one-year consensus summary into an error: the summary horizon and the underlying report's horizon are separate pieces of metadata.
The model-label field was populated in every matched record. In contrast, the selected per-report formula-rating, formula-explanation and subjective-confidence fields were null in every record. The original question was whether model and formula labels agreed. There were zero comparable populated pairs, so an agreement percentage would have no valid denominator.
An empty per-report formula field does not show that the separately stored consensus calculation is missing. Likewise, an empty subjective-confidence field is not a confidence of zero and does not tell us whether a confidence statement existed elsewhere in an original report. The finding concerns these selected stored fields, not every retained document or the full execution process.
This evidence changes my inspection procedure: preserve the model judgment and the aggregate consensus signal as separate outputs. Ask where each came from and whether a claimed comparison actually has the necessary records. Our AI-ranked stocks guide studies score and shortlist interpretation; this guide examines the construction and identity of the rating layers.
A score still needs a mapping rule
A continuous score becomes a categorical rating through boundaries, filters or additional guards. Moving a boundary can change the label without changing the underlying research. A return floor can also prevent a favorable-looking numerical score from becoming a favorable categorical signal.
For any provider, record both the number and its mapping rule. Ask whether the bands are absolute, relative to the covered universe, or conditioned on other measures. Inspect what happens at the boundary and which guards can override the initial category. Do not assume “neutral” always means zero expected growth.
Keep rating construction separate from portfolio construction too. A collection of highly rated companies does not specify weights, concentration limits, execution prices or a review schedule. Those are additional decisions. A provider's rating cannot answer them merely by using a confident label.
When a rating changes, compare the preserved editions. Was there new company evidence, a changed market price, a different peer population, revised weights or a new model? Each can move the output. A change alone does not identify its cause.
Check validation separately from the scoring recipe
A formula can be reproduced precisely and still fail to predict useful outcomes. Reproducibility tells you how the output was created. Validation asks whether it works for its stated objective under a suitable test. Keep those claims separate.
For a probability model, ask whether predicted probabilities match observed frequencies over the intended horizon. For a ranking, ask how its ordering relates to a prespecified outcome. For a portfolio strategy, require weights, costs, timing and an appropriate comparison. Historical return figures without those definitions do not establish the same proposition.
The archived label counts and hypothetical calculation above supply no such performance test. They establish record structure, a missing comparison layer and an arithmetic illustration. The model labels may share inputs, model behavior or assumptions. The number of contributors does not measure independent confirmation.
Keep a calculation receipt before relying on a rating
My practical receipt has seven items: provider, target outcome, covered universe, evidence date, horizon, calculation version and score-to-label mapping. Add the relevant source location and the specific assumption you could not verify. This makes the next research step concrete.
Start with one rating and locate its inputs. Recalculate any disclosed arithmetic. Identify which displayed diagnostics enter the equation. Then check the category boundary and preserve the edition. Mark unavailable fields as unavailable rather than filling them with an attractive interpretation.
If the output comes from a generated thesis, trace a consequential claim to its supporting document. If it comes from a trained predictor, inspect its target and validation evidence. If it comes from a deterministic formula, inspect units, weights and guards. Our investment research transparency guide provides a wider framework for preserving that evidence trail.
How are AI stock ratings calculated? The useful answer is a documented chain from evidence to transformed inputs, model or formula, score and label, with missing information kept visible. A rating becomes easier to evaluate when you can state exactly what it measures and which part of its construction you have actually checked.
Evidence and source dates
First-party extraction on 12 October 2026: 766 original Batch 7 equity consensus records, 383 companies and 5,362 unique matched contributing rating records. Equal-voice mode; original snapshots; latest active rating metadata version two. Price anchors: 18 September; contributor snapshots: 20–21 September; scoring: 21 September 2026. All joins and category counts were checked. The missing-field result is limited to the selected per-report fields. Exact SQL, input records and reproducible arithmetic are retained with this edition.
Public first-party reference: iPulse AI Consensus Score methodology, reopened 12 October. The hypothetical example uses verified stored configuration values but does not reconstruct an actual company's complete score.
Official provider sources: Seeking Alpha Quant Ratings guide and Danelfin score calculation guide, reviewed 12 October 2026. Descriptions explain methods; no independent account testing, calibration or return comparison is claimed. The cover is an original generated editorial metaphor for inputs and weighting, not market data.
