How Talakay Researches a Stock

Every analysis of a company answers two different questions, because a portfolio manager sizing a position into a print and an analyst deciding whether a business is worth owning for years are not asking the same thing.

• Short-term outlook. A view on the next earnings report: which way the committee leans (beat, meet or miss against consensus estimates), the confidence it reports in that lean, and how that compares with what the options market is pricing.
• Long-term viability. A 12–36 month research thesis: business viability, competitive moat, structural bull and bear cases, and the catalysts and risks that would change the picture.

Both come out of the same run and the same evidence, and both stay open to questioning afterwards. Neither is a recommendation to buy or sell.

A single model prompt tends to find evidence for whatever it decided first. The platform instead works the way an investment committee does, with specialists who are each asked to argue one side well:

• The Bull agent builds the strongest honest case for the name: demand signals, pricing power, the company's own history against estimates, and upcoming catalysts.
• The Bear agent argues the other side: execution risk, margin pressure, competitive headwinds, inventory and channel build, and guidance risk.
• The Quant agent reads the market's own forecast: options pricing, the historical distribution of surprises, and volatility.

The three build their cases independently and in parallel. Bull and Bear then enter a rebuttal round, where each reads the other's case and answers it directly. Finally a Consensus agent weighs what survived, and has to justify the weight it gave each side.

Your own research is an optional fourth input. Notes, channel checks or a house view go into the debate as a case the agents must engage with, and the rebuttal round pushes back on it the same way it pushes back on Bull and Bear.

Evidence is fetched when the analysis runs, not taken from a stored snapshot, and it is cited back in the output so a claim can be traced to its source.

• Fundamentals and price data from Yahoo Finance, plus the company's history of results against estimates.
• Filings from SEC EDGAR (10-K, 10-Q, 8-K) and XBRL company facts.
• News and sentiment from headline and sentiment providers.
• Options analytics from the options chain: the implied move from the at-the-money straddle, max pain and skew.

Options data is labelled as live or last-close depending on market hours, and the Quant agent treats the two differently: a stale chain is weaker evidence than a live one. You can also connect your own data-provider and model keys in settings, and runs on your account use those instead of the shared defaults.

An earnings report is judged against expectations, not in absolute terms. Sell-side analysts publish estimates for revenue and earnings per share, and the average of those is the consensus a company beats, meets or misses.

Management teams know this and often engage in expectations management: guiding estimates a little lower ahead of a report sets a bar that is easier to clear. A beat against a deliberately low bar tells you less than it appears to.

That is why the agents compare consensus estimates with the options-implied move. If options are pricing a large move while estimates look conservative, a modest beat may already be in the price, and a company can beat and still trade down on the report.

The short-term outlook is reported as an agent lean with a self-reported confidence. Both are model output. They can change with a different model or with research you add, which is why the reasoning behind them is presented first and the lean second.

The thesis looks 12–36 months out and is about the business rather than the quarter. It is built alongside the debate from the same company data and news, and the Consensus agent reads it as long-horizon context when it forms its view.

• Business viability and competitive moat: whether the model holds up and what protects it.
• Structural bull and bear cases that do not depend on one report.
• Catalysts and risks, each tied to the evidence behind it and how much weight that evidence carries.

Theses are versioned, so a new one can be compared with the last to see what changed and why.

The same analysis runs on a single company or across a book. Upload a holdings file and its equity positions can be sent into a single debate or a batch run, with each name matched to its next report date from the earnings calendar.

After any analysis you can keep questioning the Consensus agent: challenge an assumption, ask which evidence moved the confidence, or argue that a factor was weighted wrongly. It answers from the context it decided on.

The short-term outlook is the part of an analysis that can be checked against a fact within weeks, so it is the part that is graded. The lean and its confidence are recorded before the company reports, and a scheduled job scores them against the actual result using the Brier score.

Lower is better. Because it scores the confidence and not just the direction, being loudly wrong costs more than being unsure:

• 0.0 (best): a BEAT lean at 100% confidence, and the company beat.
• 0.25 (coin flip): 50% confidence, whichever way it lands.
• 1.0 (worst): a BEAT lean at 100% confidence, and the company missed.

Averaged over many reports, the score shows whether the stated confidence means anything: a system that says 70% should be right about seven times in ten. Wrong calls stay on the record.

The long-term thesis is not scored. A 12–36 month view cannot be graded on the timescale of a quarter, so it should be read as structured, sourced argument rather than a forecast with a track record.

Everything here is model output built on public data. It can be wrong, incomplete or out of date, and it is a research aid, not investment advice or a recommendation to buy or sell any security.