We measured seven kinds of corporate events. Only three move the stock, and 42% of the 'news' is noise

For two months we classified into seven kinds of corporate event every catalyst-bearing news item that entered our system: 7,046, across 365 liquid US names. Then we measured whether each kind was followed by more movement than usual in the stock. Only three are: quarterly earnings, executive changes and guidance revisions. The other four are not. And one of them — analyst upgrades and downgrades — accounts for 42% of everything the classifier tagged as relevant news. The most abundant kind of financial news is, statistically, noise.

It is worth being precise about what we measured, because the headline invites reading more into it than there is. We did not measure whether the stock goes up or down. We measured dispersion: whether, after the event, the stock moves more than it had been moving. For every item, t0 is the first session strictly after its publication timestamp — a 20:30 UTC item is already post-close and its reaction is the next session, which removes lookahead by construction — and we divide realised volatility over the following five sessions by the name's own prior volatility, estimated with an exponentially weighted moving average (EWMA, λ=0.94). A ratio of 1.0 means, literally, 'the event tells you nothing that prior volatility did not already tell you'.

That ratio is compared against an empirical baseline: the same symbols on days with no news within ±5 sessions, 1,074 observations with a mean ratio of 0.964. Significance comes from a bootstrap clustered by event date, 10,000 resamples, because items published on the same day are not independent and treating them as if they were inflates any p-value. And since we tested seven kinds, the threshold is deflated with Benjamini-Hochberg at 10%. Before running anything we wrote down a sanity check: quarterly earnings HAD to come out clearly elevated, because it is the strong, known case in the whole literature; if the method could not see it, the method was broken, not the data.

The numbers, over 5,211 events with an evaluable price window. Earnings: **1.089** (n=1,449, p<0.0001). Executives: **1.115** (n=267, p<0.0001). Guidance: **1.154** (n=52, p=0.014). All three survive deflation. Capital operations: 0.962 (n=53). Regulatory: 0.920 (n=217). Mergers and acquisitions: 0.878 (n=201). Analyst ratings: 0.914 (n=2,972). None of the last four separates from the baseline on the upside. The sanity check passed, and the magnitudes are modest: 9% to 15% more movement than prior volatility already anticipated. Useful for sizing and for excluding; not a dramatic effect, and we will not sell it as one.

Two things came out that we were not looking for. The first is ratings: 2,972 of 7,046 items, and a ratio below the baseline. A bank raising or cutting a price target is the most frequent content in the financial flow, and it is not followed by more movement than normal. The second is counterintuitive: mergers and acquisitions REDUCE volatility. 0.878, and the test in the opposite direction would give p≈0.004. It makes economic sense: once an offer is on the table, the target's stock anchors to the offer price and stops moving with the market.

The obvious objection is that two months of 2026 are a single regime. To attack it we used something we already had: 31,667 quarterly filings to the SEC with their acceptance timestamp, the legal moment the information became public. It is a different and weaker event — the filing arrives days after the earnings press release, which is the primary event — so if the effect still shows up, the phenomenon is robust. It showed up: 25,399 evaluable events, 681 symbols, 2016-2026, a ratio of 1.0066 against a baseline of 0.9240, p=0.0001. And positive in all eleven years: under the hypothesis of no effect, eleven positive signs in a row have probability 2⁻¹¹, about 0.0005. The effect also scales with market stress: largest in 2020 (+0.21) and 2022 (+0.15), nearly nil in 2018 (+0.002). It warns loudest exactly when being wrong hurts most.

Now the part we most want to publish, because it is the part almost nobody publishes. Moving from aggregate statistics to a name-by-name map exposed a problem: the data provider tags a single news item with several symbols, and our classifier matched even when the event belonged to another company. An item about OpenAI's chief financial officer showed up as an 'executive event' for Microsoft; remarks by the chief executive of an air-taxi company were attributed to Boeing. The multipliers above do not lie — they were measured with that noise inside and are its average — but a row in the map that states something false about a specific company is wrong.

We designed a filter: keep only items whose headline actually mentions the company. And before running it we wrote down the criterion for adopting it as a signal improvement: the filtered subgroup had to print a ratio at least 0.02 higher, on earnings AND on executives. On executives the filter worked comfortably: 1.131 with a mention against 1.036 without, which also stops being significant. On earnings it did not: 1.089 against 1.080, an improvement of **0.009**. **The criterion failed.** And we honoured it: the filter is NOT sold as a signal improvement, the published multipliers remain the original ones — 1.089, 1.115 and 1.154, even though the filtered subgroups measure higher — and the filter is applied only for a different, declared reason: attribution correctness. That is a product decision, not a statistical result, and so it is documented separately. It would have been easy to move the threshold after the fact, or to report only the half that worked. Doing that once is exactly what turns a study into marketing.

The A/B taught us one more thing: on earnings, the 'noise' was not noise. A competitor's earnings release moving the name too is consistent with how a sector works, and it counts for the hypothesis, not against it.

What this is not. It is not direction: our search for directional signals is closed, with an argument — ten signals re-measured with rank IC, zero survivors — and nothing above reopens it. It is not a per-name prediction: the multiplier is the average of a cohort and of a five-session window, not how much a given stock will move today. And it is not tradable: if the map says a name with earnings will move 9% more than usual, the options market already has it in the price. Nobody makes money buying volatility off this. The value is in filtering and controlling, not in betting.

What we use it for. First, a daily map generated every morning before the US open — by construction it cannot look into the future — listing the names with an active or scheduled event and the measured multiplier for its kind. Second, a 'do not trade' list: a control variable for measuring whether a guru's call was skill or an earnings calendar anyone could have read. Third, quality control on our own track record: when a net asset value series jumps, the alert arrives with the event that explains it, so we can tell an accounting bug from real news.

We publish the weekly calendar in the open: the S&P 500 names with an event in the next seven days and the measured multiplier for that kind, with the explicit label that it measures dispersion, not direction. It lives at soberquant.com/events and it is free. Every number in this text comes from reproducible scripts, and the detail — intervals, deflation and the failed criterion — is in the project documentation. None of this is investment advice; our strategies remain in simulation (paper).