The Newswire Null

Each test asks one question: after subtracting the stock’s own drift and the market’s move, does this category of headline predict which way the price goes next? Every answer came back no.

Analyst upgrades, earnings beats, guidance changes, dilutive offerings — these are the headlines with genuine information in them, and they are unambiguous about direction. The five-year sample is large enough to have caught a real effect in most of them. The detectable column is the smallest edge each test had an 80% chance of finding.

Analyst upgrades are the clearest case. Across 967 of them, a 26 basis point edge would have shown up. The measurement came back at three. That is not a sample too small to see anything — it is a good look at nothing.

Reported on its own, stripped of the other 119 tests it was found among, this is a strategy with a perfect record. That framing is not a lie about the arithmetic. It is just a lie about where the number came from.

A round trip costs about nine basis points in spread and slippage on liquid names. Every measured edge here sits at or below that line, so even taking the numbers at face value there is nothing left after paying to trade them.

The overlap column is the share of events whose measurement windows share price bars with a neighbour. At 86–92%, these are not independent observations, which means the p-values are already flattering. The null is stronger than it looks.

A null on direction is not a null on everything. Three patterns in this data are real, and none of them is an edge.

Whatever a headline does to a price, it has finished doing it well before the close. At two and three days out the figures are 0.995× and 1.000× — indistinguishable from an ordinary Tuesday.

Earnings and guidance arrive during genuinely turbulent stretches. Lawsuits, FDA approvals, buybacks and acquisitions are followed by days quieter than average — a fifth below baseline, at vanishing q-values.

The tempting reading is that boring news makes for boring days. It is more likely backwards. A newswire covers large caps continuously, so administrative filler gets written because nothing else is happening. The category is describing the day it was published into, not forecasting the one that follows. With publisher timestamps that confound cannot be separated out here.

If a wire writes because a price moved, then most headlines are stale by construction and the null above is an artifact of measuring commentary. That is the strongest argument against everything on this page, so it gets tested directly: sort every headline by how long the stock had been quiet before it arrived.

The gradient is clean and monotonic. The less time since the last story, the more the price has already done. Reverse causality, measured rather than asserted — and it is most of the corpus: only 1.4% of headlines are first mentions, while seventy percent arrive within four hours of another.

This matters more than another null, because it removes the best remaining explanation for all of them. The clean 1.4% behaves exactly like the polluted 98.6%.

The honest limit: 549 events can detect a 26 basis point edge at 80% power, so something between 10 and 25 basis points could still be hiding — tradable after costs, invisible at this sample size. Sharpening it means a wider universe, because the first-mention filter throws away almost everything.

An average of zero can still hide a payoff worth having. If news made enormous gains likelier without making enormous losses likelier, a portfolio of small positions would profit from the asymmetry no matter what the mean said. So: how often does each tail actually arrive?

News genuinely does fatten the tails, and the effect strengthens the further out you go — from 1.15× at five percent to 1.35× at twenty. That is a real finding, and it is the first thing in this study that grows rather than vanishes under scrutiny.

It is also useless, because the downside fattens by the same amount. At every threshold the negative ratio is equal to or larger than the positive one. Measured skew after news is +0.65 against a baseline of +0.74: news events are marginally less right-leaning than ordinary days. You are not buying a lottery ticket. You are buying a wider distribution, in both directions at once.

And the structure itself deserves a plain look. Over ten days, a 2% stop paired with a 20% target is hit — the stop, not the target — 71.1% of the time. The target arrives 3.6% of the time. A 2% stop on a stock that routinely moves 2% in a day is not risk control; it is a near-guarantee of being removed from the position before anything can happen. Ties were resolved in the strategy’s favour nowhere: when both barriers were breached in one session, the stop was counted first.

Everything above is a null. This is not. News filed before the opening bell produces a gap, and then the price continues in the same direction through the session — the opening auction underreacts.

The effect is symmetric, which rules out the obvious artifact: positive news gaps up then drifts up a further 23.6 bps, negative news gaps down then drifts down 21.6 bps. A constant intraday bias cannot produce both.

7,781 trades over 1,202 trading days. Per trade: mean +22.8 bps, median +13.6 bps, win rate 54.0%, best +25.1%, worst −19.8%.

Read the two rows against each other. Eleven basis points of assumed cost is the difference between 37.8% a year and 7.0% a year — and at 7.0% with a 31% drawdown you have taken considerably more risk than an index fund to earn less. The entire result lives inside the execution assumption, and it must be executed at the opening auction, the widest-spread moment of the trading day.

Entry at the open and exit at the close is a day trade. Below $25,000 in equity that is capped at three per five rolling days, against a median of six signals a day. Two thousand simulations of an account taking three a week at 25% position size:

A null is only interesting if you can say what produced it. These four steps are ordered, because each one removes part of whatever edge the previous step left standing. By the fourth there is nothing to remove.

None of this makes markets mysterious or the study a failure. It is what an efficiently priced market is supposed to look like from the outside: information that is genuinely valuable, distributed to everyone simultaneously, absorbed faster than a person can read it, leaving a residue too small to pay for its own collection.

The finding is not that news does not matter. News moves prices — that is visible throughout these results. The finding is that by the time you can act on a public headline, the moving is done.

This is a measurement of historical data. It is not investment advice, a forecast, or a recommendation about any security.