Note
Grounded in the stored record and live market data. Not investment advice.
loading…
The record is organised in three layers. The macroeconomic layer holds rate, inflation, and employment releases together with central bank policy statements. The microeconomic layer holds company-level events: regulatory filings, reported and scheduled earnings, and company news. The markets layer holds index-level and cross-sector news. Filings and macroeconomic series extend back five years; news extends to March 2022. Records accumulate rather than being replaced, so the depth of the archive increases over time.
Sixty-three FRED series are read, forty-six macro and seventeen markets.
Rates. Treasury yields at three months, two, five, ten and thirty years, the ten-year minus two-year curve, effective fed funds, SOFR, the bank prime rate, the thirty-year mortgage rate.
Inflation and labour. CPI, PCE, breakevens at five and ten years, unemployment, nonfarm payrolls, weekly jobless claims, job openings, average hourly earnings, participation.
Activity. Housing starts, retail sales, industrial production, durable goods, the saving rate, Case-Shiller, vehicle sales, consumer sentiment, GDP.
Markets. The S&P 500, Nasdaq 100 and Dow, two volatility indices, WTI and natural gas, gold and copper, five currency pairs, the dollar index, investment-grade and high-yield spreads.
Sovereigns. Ten-year government yields for around forty economies across North America, Europe, Asia-Pacific and Latin America. Short rates and unemployment are read for the core five, the United States, Canada, Japan, the United Kingdom and Germany; for the rest only the yield and the exchange rate are available on free data. China is covered by its exchange rate alone.
A release becomes an event on publication. A continuous series becomes an event only when it moves, and the bar is set per series rather than shared.
Each threshold is two standard deviations of that series' own daily change distribution, measured from 2022 to 2026. A one per cent move in the dollar index and a one per cent move in crude are not comparable events, so they do not share a number.
At two sigma a daily series fires nine to fifteen times a year. The ten-year Treasury yield has a daily standard deviation of 6.2 basis points, so its threshold is 12.4 and it clears on about five per cent of sessions. At one sigma it would clear on a third of them, which is the market's ordinary churn rather than news.
Fourteen market series also record a weekly close, so the lane stays continuous through quiet stretches rather than reading as missing data.
Recording that a series moved is not useful on its own. Anyone watching the screen saw it. The question that is worth answering is what has usually followed.
For any series and threshold, every historical breach is located and the following sessions are measured, both in that series and across the others. The result is a distribution with a sample size attached rather than a single number.
The first finding is a negative one. After a two-sigma move in the ten-year, the next session is up 51% of the time across 63 breaches, and the next five sessions are the same. There is no reliable short-horizon follow-through. The cross-asset response is more directional: the curve flattens in 60% of cases and volatility falls in 59%.
Sample sizes are reported on every result and anything below ten observations is marked unreliable. Forward windows from clustered breaches share sessions, so the dispersion shown is a floor rather than an estimate.
The coverage universe comprises the S&P 500 constituents, twenty-five cross-listed Canadian issuers, and companies admitted through repeated co-occurrence with covered names. It also holds sovereign coverage for around forty economies across the major advanced and emerging markets, and commodity coverage for crude oil, natural gas, gold and copper. Filings are matched by SEC Central Index Key, which is an exact identifier. Prose is matched by company name, and ticker symbols are matched case-sensitively, which prevents ordinary English words from being read as symbols.
Beyond company stocks, a country and a commodity are each a coverage entity with its own brief, selected the same way a stock is. A sovereign brief reads the government bond curve, the policy rate, the exchange rate, and the rate differential to the United States, which for an issuer that borrows in its own currency is a policy and inflation gap rather than a credit signal. A commodity brief reads the price level and its momentum against its own history.
Each brief states a base, bull and bear case with the trigger for each, names one non-obvious second-order consequence and its mechanism, interrogates the intent behind any policy signal or leak by asking who benefits and why it surfaced now, and offers the closest historical analogue marked as measured or remembered. Every driver either cites a numbered event or is tagged an assumption. The futures curve, inventory, and true sovereign credit spreads are stated as unavailable rather than invented.
Three lanes separate the macroeconomic, markets, and microeconomic layers. Each marker represents a single event and is scaled by assessed importance. Price and market capitalisation for the selected company are plotted against the index, with each series indexed to 100 at the start of the window so that relative performance is directly comparable.
The window is moved by the range presets, by dragging the timeline, by scrolling to zoom on the cursor, or by dragging the overview strip beneath the toolbar; the range presets highlight only while the window matches one. Filter by category, ticker, or event type. A stock rating is generated on demand by checking Stock Rating for a selected name, and clicking a rate or macro-series event appends what has historically followed a move of that size. Citations beneath an answer link to the underlying event.
The vector index holds approximately 6,500 entries against more than 70,000 stored events, which limits semantic search to under a tenth of the record. Full-text search covers the entire corpus at no per-query cost. Lexical retrieval therefore serves as the primary method, with the vector layer providing an additional signal over recent events.
Queries requiring counts, rankings, or comparisons are computed in SQL across the complete record rather than estimated from a retrieved sample. Such figures are exact, and the response identifies them as computed. A retrieved subset is never reported as a total.
A score is derived from five years of financial statements before any news is considered. Valuation, the estimate record, filings, and analyst consensus subsequently adjust that score within a fixed bound, so that recent headlines cannot override the multi-year financial record. Fixed thresholds convert the score into a Buy, Hold, or Sell classification, and the score governs that classification even where the written argument names a different word. Where operating cash flow has trailed net income for two consecutive years, the score is capped. A rating is produced only on request, and the window it judges is stated on the note.
Margins, returns, and growth are compared against the median of covered peers within the same GICS sector, and the peer count is reported alongside the comparison. A gross margin above the sector median indicates pricing power. A return on equity above the median that depends on higher leverage is identified as such rather than reported as an advantage. Where fewer than three peers have been collected, no sector comparison is drawn.
Every figure in a note must appear in the material supplied to the model. Peer multiples, price targets, and sector averages are not generated where none were provided, and their absence is stated. The counter-argument is constructed from the weakest figure in the supplied data, which is identified independently of the model.
An event may remain operative long after it occurs. Rate cycles, restructurings, and repurchase programmes affect reported results for several years and appear as level shifts in margin, cash conversion, or share count. An interval containing no events is treated as ordinary and does not reduce the score.
Equity. The system does not access earnings call transcripts, analyst price targets, or private company records, each paywalled or absent from public filings. It offers no assessment of supplier power, buyer power, or substitution, since the data does not support one.
Sovereign. Coverage spans around forty economies, not the whole world, and for most of them only the ten-year yield and the exchange rate are read; short rate, inflation and unemployment are available for the core five only. Free data reaches roughly forty mostly-OECD economies, and frontier markets have no free sovereign yields. China is exchange-rate only, because no free ten-year series exists for it. Outside the United States the yield, rate, and unemployment series are monthly rather than daily, only a ten-year and a short rate are read rather than a full curve, and inflation is not yet wired. The gap to the United States ten-year is a rate differential, not a credit spread: for an issuer that borrows in its own currency it reflects policy, not default risk, and a true sovereign credit read, whether CDS or emerging-market EMBI spreads, is not yet in the data.
Commodity. Only the spot price, its momentum, and its percentile are read. The futures curve and inventory are not available, so contango, backwardation, and the storage picture cannot be assessed. Gold is daily; copper is monthly.
Reasoning. The written view is generated by a language model. Every figure is copied from the supplied data, but the interpretation is the model's and can err. A driver must cite a numbered event or be tagged an assumption, though the citation is enforced by instruction rather than verified afterward. A historical analogue older than the record, which begins in 2022, comes from the model's general knowledge and is marked remembered rather than measured. Event retrieval is keyword-based and can surface an off-topic match.
Data. The vector index holds about 6,500 of more than 70,000 events, so semantic search reaches under a tenth of the record while full-text search covers all of it. The news archive begins in March 2022; filings and series reach five to ten years. Output is not investment advice.
Collection runs every fifteen minutes on a scheduled trigger, independently of page requests.
How a threshold crossing becomes something with a sample size behind it.
Breaches that cluster produce overlapping forward windows, so the observations are not independent and the spread understates the true dispersion. This is stated on every result rather than left for the reader to discover.
Only the model call runs on request; all other inputs are read from storage. Latency or failure in an upstream provider therefore cannot interrupt a note.
If the model is unavailable, the most recent saved note is shown with its original date. The accompanying figures remain current.
Four retrieval methods run in parallel and their results are merged.
Companies appearing alongside covered names at least three times are resolved and added to the universe.
Symbols that do not resolve to a listed security are not added. They are re-checked on subsequent appearances.
Grounded in the stored record and live market data. Not investment advice.
Answers are grounded in the stored events, with a citation for each one used.