Methodology
How each part of Dawo produces its numbers, and how each is measured.
How a read is made
Open a name and Dawo shows what today's price needs from the drivers the company itself reports, against their own trend, and where the evidence leans. There is no verdict on that page, no target and no conviction score. The read is built in four steps, and the split between what a model does and what arithmetic does is the point.
1. The drivers are the ones the company states
A model reads the last eight earnings releases and lists every operating metric the company gives with a number: data-center revenue, paid seats, price per ad, gross margin. It then names a structure for the company — which reported segments there are, which stated lines sit inside each, how they combine — using only that list. It may not name a unit, a customer count or a price the company does not disclose. The structure must reproduce reported revenue from the segment table within 2% in at least two fiscal years, or it is rejected.
2. Every number is a print
The figures in the table come from three places, in order of trust: the XBRL statements the company files, the segment tables in its annual report, and the sentences in its releases and filings. A smaller model copies the sentence that states a figure; the figure is admitted only if the sentence is verbatim in the document, contains that exact number, is not guidance, and describes the driver it is filed under. A quarter-over-quarter figure is refused for a growth driver. Shares of a segment come from the annual report's product table or a stated dollar level, never from the model. No model writes a number that reaches the page.
3. The arithmetic is the engine's, and you can open it
Each driver's trend is its own history, bounded by the same growth range the valuation model applies. "Price needs" solves one driver at a time for the value that makes the model equal today's price with the others at trend, and says so when a line cannot reach it alone. The headline figure is the single growth delta every priced driver would have to run against trend for the price to be right. Claims from the filings and calls are sorted to the driver they bear on and shown as a lean with counts for and against, never as a score.
4. A person reviews the structure, and your own calls are kept
A structure serves as a draft until a reviewer checks the segments, the lines, the shares and a sample of the claim tags; the reviewer's name is on the page. When you type your own rates and record what you believe, Dawo keeps that with the values you set and the price they reprice to, watches the next print against the condition you wrote, and shows you how your calls have done. That record is yours, and it is the only view on the page.
The engine's forward track
Alongside the read, the engine runs a full analyst process on each name and records a view. Those views are not shown on a symbol page; they are kept so the engine's assumptions can be measured against what happened. This is how that process works and how it is scored.
How Dawo analyses a company
Every company gets the same analysis: the same stages, in the same order, whatever the company. The valuation itself is arithmetic, not a language model — the discounted cash flow, the multiples, the peer regression and the forensic-accounting checks are ordinary code, and they produce the same number twice from the same inputs.
Language models do the parts that are genuinely reading and judgement, and more than one is involved because the jobs are not alike. Pulling the risks, guidance and commitments out of a filing or an earnings call is careful extraction, and runs on smaller, faster models. Weighing what all of it means — choosing the assumptions, arguing the bear case against the bull case, and settling on a view — is the hard part, and runs on the most capable model available. We do not publish which model does which, because that changes as better ones ship and a page naming them would be out of date before it was useful.
The output is an Undervalued / Fairly valued / Overvalued assessment with a conviction score, bear, base and bull price scenarios, and a stated view on what the market may be missing. It is an assessment of value, not an instruction to buy or sell.
The pipeline
A six-stage sequence. Each stage has defined inputs and outputs; stages are not skipped.
| # | Stage | What it does |
|---|---|---|
| 1 | Understand the symbol | Declare the business model, the primary value drivers, and which valuation methods fit this company — and which will mislead, and why. The framework is set before any numbers are run. |
| 2 | Build scenarios | Bear, base, and bull cases, each with a named catalyst and its mechanism (e.g. a specific approval, a margin inflection, a segment ramp). |
| 3 | Value each scenario | A custom DCF plus forward P/E and peer multiples for each scenario. A reverse DCF back-solves the growth the market is currently pricing; a sum-of-parts segment model is used when the company reports segments. |
| 4 | Stress test | Scenario stress across the bear/base/bull bundle, with an optional per-assumption tornado ranking. Required before the report can be finalized. |
| 5 | Cross-signal check, challenge, and score | Insider activity, forensic-accounting flags, momentum, short interest, and options positioning read together, to surface patterns a single signal would miss. Then an adversarial bear challenge AND a bull champion (both required); the drivers that move the value, each named against the model input it acts through; and a six-dimension scorecard — valuation, quality, momentum, consensus, risk, optionality — whose vote count sets the suggested view. |
| 6 | Finalize | The final assessment must match the scorecard unless overridden with a stated reason, and must say what the market appears to be missing — tied to named catalysts — along with a risk/reward ratio and how long the catalyst is expected to take. |
Coverage and refresh
- The same pipeline runs for every symbol — there is no tiered analysis depth.
- Re-analysis is event-driven: a re-run is triggered by an earnings report, a price move of ≥10%, a new SEC filing (10-K / 10-Q / 8-K), an estimate revision, or an analyst rating change.
- Inputs are point-in-time: SEC filings, FMP consensus estimates and segments, price history, and earnings transcripts, extracted into forward signals before the analyst runs.
- Research, not advice: the assessment is not personalized and is not a recommendation to buy or sell.
The forward record is building
Every recorded view is timestamped and scored against actual market and S&P 500 returns over its 3, 6 and 12-month horizon, with hit thresholds fixed in advance. Nothing is published as accuracy until a horizon has enough views to mean something; the first are expected in Q4 2026. What is shown here is the count so far, and the misses will be shown beside the hits. The product's own record, what each read said the price needed and what the company printed next, is the ledger. Both are reachable from Claude or ChatGPT through the connector.