Which way do alternative-data signals point?
Insider buying is bullish, bad headlines are bearish, an expensive borrow is a warning. Everyone builds on those assumptions. We measured them instead, and two of the three turned out to be backwards.
The question
Every strategy that consumes an alternative-data signal assumes a direction for it. Those assumptions are usually inherited rather than tested, and a risk overlay that only trims positions can never reveal that its own premise is wrong. So this study asks the prior question: for insider flow, news sentiment and borrow fees, which way does each signal actually point?
The instrument is deliberately dumb. For each signal, each stock goes long when the signal sits in the top quartile of that stock's own history, short in the bottom quartile, flat in between. The rule is symmetric and builds in no view: if the assumed direction is right, the book makes money; if it's backwards, the book loses. We ran it on two ten-name panels (mega-cap tech, and moderate-beta banks and industrials) at three holding periods, plus a variant restricted to uptrending names: 36 backtests over 2019–2026.
Two textbook arrows, reversed
Insider extremes are bad news in both directions. Books long insider-buying extremes and short insider-selling extremes lost money in 10 of 12 configurations, worst on tech at −2.4 to −5.4% a year. Combined with earlier findings, the pattern looks non-monotonic: unusually heavy insider activity in either direction precedes weaker returns, and heavy buying is the worst of it.
Expensive borrows outperform. Stocks with spiking borrow fees beat cheap-to-borrow ones in 9 of 12 configurations: small, but consistently opposite the standard warning-flag reading. These two reversals explain, in hindsight, why overlays built on the textbook directions never worked.
The one survivor and its fair trial
News sentiment pointed the assumed direction: positive in all six trend-restricted configurations, around 8% a year on the tech panel, with near-zero market beta. It passed every mechanical check we could throw at it: no execution look-ahead, both legs active, market exposure removed.
But two flags stood out. Removing a single stock (NVDA) took the result from significant to not. And it was the best cell among six signal-variant combinations, exactly where a lucky result would sit. So before believing it, we froze the spec exactly as discovered and pre-registered a confirmation with a mechanical pass/fail rule: ten fresh tech names over the same period, and the original names over 2015–2018, a period untouched by discovery.
The rule, fixed before any confirmation run: confirmed required positive means in both tests and pooled p ≤ 0.05; pooled p ≥ 0.32 refuted. Had the true effect matched the discovery estimate, the expected pooled p was about 0.02, so the confirmation had adequate power and failed anyway.
Where this leaves us
Across three signals, two panels and 58 backtests: these signals carry some directional information, enough to show two textbook assumptions are wrong at extremes, but the effects are far too small to trade. The general news effect measured +0.6 to +1.3% a year, an order of magnitude below the discovery number.
The confirmation protocol is the part we'd highlight to an allocator. The pass/fail rule was written down before the runs; the verdict was reported as computed; the study ended there. One question remained open: do these conclusions survive at institutional breadth? Note 03 answers it.
Panels: tech (AAPL, MSFT, NVDA, AMZN, META, GOOGL, TSLA, AMD, NFLX, PLTR) and moderate-beta (JPM, BAC, GS, CAT, BA, DIS, GE, UBER, QCOM, DAL); confirmation panel CRM, ORCL, ADBE, INTC, CSCO, TXN, MU, AVGO, NOW, IBM. Data: Quiver insider transactions, Tiingo news sentiment, Interactive Brokers borrow fees, point-in-time throughout. Signals scored against each stock's own one-year history. Significance from weekly returns with Newey–West standard errors; alpha vs SPY; leave-one-out re-runs per name; fill-log checks against look-ahead. The books are built to measure direction, so borrow costs are not modeled. Backtested results are hypothetical; this note is research and is not investment advice.