Earnings Preview

Earnings Preview

Everything in this note locates the stock against one object: the bar, the level of results the market is already positioned for. Consensus is one input to the bar, not the bar itself. The preview says where the bar sits, what could clear or miss it, what to listen for on the call, and what the reaction is likely to be if the print lands each way.

After the print, route to .agents/skills/earnings-analysis/SKILL.md.

Evidence labels, source tiers, staleness, the readiness posture and the intake limits: .agents/skills/research-conventions/SKILL.md, read before the first deliverable.

Route

The full preview is the default. A short version is produced only when the user asks for one, and it is rebuilt at the lower depth rather than truncated (.agents/skills/research-conventions/references/depth.md): the freeze time, the source posture, the bar numbers, the top two debates, the must-watch call questions and the missing-evidence block all survive the rebuild.

Missing inputs never shorten the note. A section whose input is unavailable keeps its heading and carries a labelled gap saying what is missing and what would supply it. A short note reads as a thin setup; a full note with three labelled gaps reads as the truth.

Correctness rules

Every number follows the shared market-data rules: .agents/skills/research-conventions/references/market-data-rules.md covers fiscal periods, per-company LTM, NTM as four quarterly estimates, one base date for comparative returns, shown arithmetic, rate against level changes, margin selection and date-stamping. Read it before the first pull.

Preview-specific, on top of those:

• Freeze the clock. Record one as-of moment for the whole set of time-sensitive inputs: consensus, whisper, options, price, positioning. State it at the top of the note. An input with no as-of is a gap, not an estimate.
• Suppress a percent change when the denominator is zero, negative or definitionally unstable, and show the absolute delta instead.
• Mark a year-over-year comparison non-comparable when the company changed the metric's definition and published no restated bridge.

Step 1: Freeze the clock and map the period

Pull the setup: get_company_overview for consensus, earnings history and price targets; get_daily_prices for the price history and the event window; get_sec_filing for the prior quarter, with its transcript attached, to recover the guidance in force; get_shares_float and get_short_data for the positioning picture; WebSearch and WebFetch for news since that filing.

Pin the period map once and reuse it everywhere, taking each fiscal label from the earnings event name get_company_overview returns: the preview quarter, the prior quarter, the year-ago quarter and the two-years-ago quarter. Every table, chart, scenario and options reference uses those same four labels. Ask the user only when the preview quarter is genuinely ambiguous, for example when the company reports two quarters inside one month.

Complete when the four period labels are written down, the freeze-time as-of is recorded, and the earnings date is confirmed from the company or get_earnings_calendar rather than inferred.

Step 2: Build the expectation bar

Five expectation sources, kept separate because they say different things. Collapsing them is how a preview ends up measuring the print against the wrong number.

Close the table with one sentence stating where the bar sits and why: which of the five the stock is trading against, and how far the others sit from it.

Whisper rules. A whisper is never blended into consensus without showing the bridge. With strong support, carry it as a value with its provenance and a confidence label. With weak support, convert it to qualitative setup language rather than a number. With no external whisper at all, either derive an implied whisper from the guide midpoint plus this management's beat history and label it analyst-derived, or write "not provided" and leave the row empty.

Complete when all five rows are filled or carry an absence reason, and the bar sentence states the selected bar's value, source, vintage and basis, labelled judgement wherever the five rows leave competing bars in play.

Step 3: Watch the EPS quality before the print

Headline EPS is a bad proxy for recurring earnings more often than a preview assumes, and the items that break it are knowable in advance.

Screen at least these: the effective tax rate against the guided or trailing rate, the diluted share count (buyback timing, issuance, convertible dilution), below-the-line and mark-to-market items, FX translation and remeasurement, disposals, impairments, restructuring and litigation.

Then check the basis: does published consensus measure the same thing the company reports? Where it does not, say so, and say which figure the bar actually rests on. The real bar is often revenue, operating income, segment profit or free cash flow rather than EPS, and a preview that assumes EPS measures the wrong quarter.

Complete when each screened item carries a pre-print risk, and the note states whether consensus and company bases match.

Step 4: Bridge the guide to an implied bar

A guide is a starting point, not an expectation. Convert it using this management's own record.

• Beat frequency over the last eight quarters, against its own guide, per metric.
• Typical beat size, in currency for level metrics and basis points for rate metrics.
• Whether the conservatism is steady, widening or fading, and what changed if it moved.
• The limits behind those statistics: how many quarters the sample has, whether the management team or the guidance policy changed inside it.

Complete when the implied bar is stated as a number or range with the beat history that produced it, and the sample limits are named.

Step 5: Choose the metrics that decide the quarter

Financial: revenue in total and by segment, EPS, gross and operating margin, free cash flow, and forward guidance against consensus. Rank them by which one moves the stock, not by which one leads the P&L.

Operational, by sector, as the default pack: software and internet (ARR, net revenue retention, remaining performance obligation, customer count), retail (comparable-store sales, traffic, basket, inventory), industrials (backlog, book-to-bill, price against volume), financials (net interest margin, credit quality, loan growth, fee income), healthcare (scripts, patient volumes, pipeline events), energy and materials (volumes, realised price, unit cost).

Selection rule. The issuer's own disclosure model beats the generic pack. Use the metrics this company guides on and gets asked about, and where an important sector metric is not disclosed, list it as a data request rather than dropping it silently.

Flag as material: a gap between guide, consensus and whisper wider than the sector norm; a metric breaking its trailing four-quarter slope; growth decelerating or margin compressing past a stated threshold in basis points. For a seasonally distorted metric or one with a distorted base year, show the two-year stacked growth rate beside the year-over-year rate.

Complete when the metric list is ranked, each entry carries the expectation it will be measured against, and undisclosed metrics appear as data requests.