Wrong number
A model supplies a plausible figure without retrieving or calculating it from evidence. The answer looks precise but cannot be reproduced.
Financial AI reliability
Financial AI should never ask you to trust a number you cannot verify.
Altys separates language from calculation. Financial figures come from sourced data, material claims carry citations, historical answers respect what was knowable at the time, and core outputs can be exported so an analyst can reproduce the work outside the software.
Calculated, not generated
Cited at the claim
No evidence, no answer
Point-in-time, not hindsight
Exportable audit trail
What hallucination means in finance
The most dangerous error usually looks reasonable.
A financial hallucination is not only a completely invented figure. It can be a real value with the wrong label, a valid disclosure used after it became stale, or a citation that sits beside a claim it does not support.
Wrong period
A quarterly number answers an annual question, or a year-to-date figure is compared with a full year. Both inputs may be real and the comparison still invalid.
Wrong entity or basis
Standalone and consolidated accounts, parent and subsidiary, or segment and group values are quietly mixed into one conclusion.
Wrong vintage
Today's corrected history is presented as if it had been available on a past decision date, introducing hindsight into research or a backtest.
The reliability standard
A citation is necessary. It is not sufficient.
A real filing can be attached to the wrong sentence. Reliable financial AI must prove that the cited evidence supports the exact company, period, reporting basis, unit and claim.
Right sourceThe evidence comes from the applicable filing, statement, presentation or transcript.
Right scopeThe company, entity, period, basis, unit and currency match the question.
Right mathDerived values can be reproduced from named inputs and explicit methods.
Right timeHistorical research uses only information that was available by the date.
The citation is a route to scrutiny, not a badge of truth. It should make checking the answer faster, not ask the analyst to suspend judgment.
The Altys control stack
Language, evidence and calculation stay separate.
The system assigns each job to the layer that can perform it reliably. A language model helps read and explain. It is not the source database, calculation engine or final investment decision.
01 · SOURCE
Start with the primary record
Company filings, XBRL statements, concalls, guidance and other research evidence enter with source and identity metadata.
02 · STRUCTURE
Pin the financial context
Company, period, consolidated or standalone basis, unit, availability date and later vintages remain explicit rather than being inferred afresh from prose.
03 · CALCULATE
Run math through code
Ratios, factors, growth and forecast outputs use explicit methods over sourced inputs. Missing data remains missing and does not silently become zero.
04 · RETRIEVE
Find evidence for the question
AI searches the appropriate company evidence and structured data instead of relying on a model's general memory of the company.
05 · CITE
Attach support to material claims
The source sits beside the claim so an analyst can inspect the actual evidence, not only a bibliography after the answer.
06 · GATE
Fail closed when evidence is missing
An unsupported answer resolves to error or unavailable. Fluency cannot promote an unverified claim into a completed research result.
Worked example
Revenue growth should be a calculation, not a sentence completion.
Consider a user asking for annual revenue growth when the source shows ₹100 crore in the prior year and ₹118 crore in the current year.
Same question · two system designs
Generated answer
“Revenue grew by about 20 per cent.”
The value may sound reasonable, but the inputs, period, basis and rounding method are invisible. Asking again may produce a different result.
Calculated answer
(₹118 crore / ₹100 crore) - 1 = 18 per cent
Both source values, their annual periods and reporting basis are attached. The result is reproducible, and the AI can explain it without originating it.
Exportable verification
Do not trust the black box. Open the workbook.
A result is easier to trust when it survives outside the product that produced it. Altys exports the working layers behind core research outputs, not only a polished final answer.
The verification surface follows the work. Formula-native Excel where calculations should recalculate, Excel-ready CSV where grids should remain portable, and versioned rules plus replay where an alert needs to explain why it fired.
Strategy screensExport the basket, full eligibility universe, factor scores, editable weights and method sheet as a formula-native Excel workbook.
ScorecardsInspect raw inputs, peer basis, derivations, scores, weights and provenance, with live spreadsheet formulas rather than pasted values.
GenGrid and analyticsDownload current columns, sort, data and persisted citations as Excel-ready CSV for independent checking and further analysis.
AlertsAudit the pinned rule definition, version, fired-by context and historical replay that explains when and why a rule would trigger.
Point-in-time reliability
The answer must respect what could have been known.
A current database can contain restatements, reclassifications and corrections published long after the original result. Altys preserves each vintage and the date it became usable, so a historical answer does not quietly borrow from the future.
Current analysisUses the latest valid record, with source and reporting basis visible.
Historical analysisUses the version that was available by the chosen date.
RestatementsImprove today's record without erasing the information set investors originally saw.
Composite metricsInherit the availability of every required input rather than pretending the result existed earlier.
What AI does and does not control
Reliability improves the decision process. It does not remove judgment.
Altys can make facts traceable, calculations reproducible and evidence faster to review. It cannot decide which assumption deserves belief, which risk matters most, when to act or how much capital to allocate.
- AI can: retrieve, compare, summarise, explain and draft from evidence.
- Code can: calculate ratios, factors, forecasts and backtests reproducibly.
- The system can: preserve history, citations, availability and failure states.
- The investor must: set assumptions, challenge the thesis and make the portfolio decision.
For a practical evaluation, read seven tests before trusting AI stock analysis or use the 60-second verification checklist.
Questions, answered
What teams ask about financial AI reliability.
Does Altys hallucinate financial numbers?
Altys is built to prevent hallucinated financial numbers. Figures come from sourced financial data and deterministic calculations, material research claims carry citations, and unsupported answers fail closed as unavailable instead of being completed with a plausible guess. No responsible system should replace analyst verification or claim that every possible error disappears.
Are Altys financial ratios and forecasts generated by AI?
No. Ratios, factors and forecast outputs run through explicit code-owned methods using identified inputs. AI can retrieve, compare and explain the result, but it is not the calculation engine that originates the number.
Can Altys calculations be exported and checked in Excel?
Yes. Strategy screens and scorecards export as formula-native Excel workbooks with inputs, eligibility, weights, calculations and provenance. GenGrid and analytical tables export as Excel-ready CSV. Alert rules are audited through versioned definitions, fired-by context and historical replay.
What happens when Altys cannot find supporting evidence?
The answer resolves to error or unavailable. Altys does not allow an unsupported material claim to pass as a completed research answer simply because a language model can write plausible prose.
Why does point-in-time data matter for financial AI?
Point-in-time data distinguishes the period a value describes from the date it became knowable. This prevents later restatements and corrections from leaking into historical research, screens or backtests.
One-month pilot
Bring a company you know. Test every number.
We will run Altys on your real research questions, source every material claim, and show exactly where the AI stops and the evidence begins.