What news does to private-company share prices

Companies that were sued or came under investigation went on to under-perform comparable periods by roughly 1% after three weeks, 3% after seven, and 6% after twelve. Nothing measurable happened in the weeks before the news broke, which is what makes the result worth reporting at all: the sequence runs news first, price second.

Ten other event types produced no effect we can distinguish from chance: funding rounds, layoffs, executive departures, down rounds, shutdowns, security breaches, product launches, partnerships, acquisitions and regulatory actions. Some look suggestive. None survive an error bar.

And the one result that does clear the bar clears it narrowly. Because we tested eleven event types, a threshold of one-in-twenty is too generous — with that many attempts, roughly one spurious result is expected. Adjusting for the number of tests, the lawsuit finding no longer reaches significance. We think it is real, for reasons given below, and we would not bet a decision on it.

Three things make a naive version of this study wrong, and we got each of them wrong before getting them right.

A mark is not a quote. A private secondary price records a completed transaction, and completion trails agreement by up to three weeks. A price printed today reflects what buyers and sellers knew weeks ago, so measuring the days immediately after an announcement compares one pre-news price against another pre-news price. Our first attempt did exactly that, found nothing on the day, and nearly concluded that private markets react slowly. They do not react slowly; the data simply cannot show a reaction there. Every window on this page is offset by that settlement lag.

These companies are going up anyway. A company that raises is a company doing well, and its mark drifts upward whether or not anything was announced. So each result is measured against the same companies at other times — fixed offsets far from any of their events. Adding that control cut the headline funding figure by more than half.

Percentage returns are useless here. Marks in this universe span three orders of magnitude over four years; one company runs from 16 cents to $237. An arithmetic mean of percentage returns produced “+93% after 60 days” from two companies, while the median sat at zero and half the events were negative. Everything here uses log returns and medians.

90,850 news articles about 130 private companies, published between 2020 and 2026, each matched to a company and judged. Matching is the part that quietly ruins studies like this: a great many private companies are named with ordinary words — Figure, Scale, Ramp, Mercury, Harvey, Kraken — so a keyword search returns mostly noise for exactly the companies people care about most.

Articles were scored by rule and, where the rules could not decide, by a model. 41,232 survived as events genuinely about the company. A further 9,455 were correctly matched and discarded as mentions: an article quoting a chief executive about the industry is about the company and is not an event, and counting those as events pulls every measured effect toward zero. Prices are daily secondary marks from 2022 onward, which is where the usable history starts.

Each figure is the median abnormal return of companies with the event, minus the median for the same companies at comparable event-free times, over a window that begins after the settlement lag. Intervals are 95%, bootstrapped over 4,000 resamples.

Only the first row excludes zero, and only before adjusting for the number of event types tested. The rest are reported because a study that publishes its one positive result and drops the others is not reporting, it is selecting.

The nulls carry information of their own. Routine good news — a launch, a partnership, an acquisition — does nothing measurable to these prices, and the point estimates lean very slightly the wrong way. Whatever moves a private mark, it is not the ordinary flow of corporate announcements.

For lawsuits and investigations the effect appears only after the settlement lag and then deepens: −1.14% at three weeks, −3.01% at seven, −5.85% at twelve. The two windows before the announcement are indistinguishable from zero. An explanation based on timing — companies being sued when their prices were already sliding — would have to produce a decline before the news as well, and it does not.

What keeps us interested despite the correction is the shape rather than the threshold. A spurious result has no reason to arrange itself: it would not appear in three consecutive post-settlement windows, deepen with each one, and leave both pre-event windows flat. That pattern is what a real effect looks like, and it is not what one expects from noise that happened to cross a line. But shape is an argument, not evidence, and the honest position is that this needs more events before anyone relies on it.

This is under-performance, not decline. Twelve weeks after being sued, these companies were up 3.00%. Comparable periods were up 8.86%. They still rose; they rose about six points less. “Prices fall after bad news” would be the wrong summary.

It cannot tell you what will happen to a particular company, and it certainly cannot tell you what to do about your own shares. These are medians across hundreds of events over four years. The spread around them is wide, and a median says nothing about any single case.

Nor is it causal. We show that prices behave differently around these dates than around other dates, and that for one event type the difference arrives after the news rather than before it. That is a necessary condition for causation and not a demonstration of it.

Known limits, in the order we would worry about them: we tested eleven event types, so the one positive result must be read against ten negative ones and does not survive a correction for that. The control periods are required to sit clear of any event for the same company, and events cluster in good times, so the control samples quieter stretches and every figure here is likely an upper bound. Three event types have fewer than fifty events. Our coverage of the busiest companies is capped by what the news index will return in a single day, so a handful of them are incomplete by an unknown amount. And headlines are classified by published pattern rules, which are legible and imperfect.

Abnormal return is a company’s log return minus the equal-weighted log return of the rest of the covered universe over the same calendar window, which removes market-wide movement. Control anchors are the same companies at fixed offsets of ±180 and ±365 days, discarded if they land within 90 days of any event for that company. Coverage of one story across many outlets is collapsed into a single event, so a widely-reported round counts once rather than a dozen times. Headlines reporting both directions at once — “raises $50M at a down round” — are dropped from both samples.

Denials, rumours, hypotheticals and questions are rejected before classification. “Company denies layoff reports” carries every keyword of a layoff and reports the opposite; counting it adds an event on a day when nothing happened.

Intervals are percentile bootstraps over 4,000 resamples of both groups, seeded for reproducibility. Groups smaller than eight are not resampled and are reported as unmeasurable rather than given a confident-looking number.