Does alternative data improve a simple trend model?
Insider filings, news sentiment and borrow fees are the most accessible alternative data there is. We gave each one a fair chance to make a trend follower better. None of them did, and the way we know that is the point.
The question
Free alternative data is everywhere: insider transaction filings, scored news headlines, the fee charged to borrow a stock. The natural hope is that these streams can sharpen an otherwise ordinary strategy: hold your winners, but hold less of the ones the data is quietly warning you about. This note tests that hope directly, in the simplest setting we could design.
Ten liquid, moderate-beta US names (banks, industrials and consumer) each run the same base rule: long when the trailing 20-day return is positive, in cash otherwise, re-checked every five trading days. Three augmented versions halve a position whenever that name's own alt-data history turns extreme: unusually heavy insider selling, unusually sour news, or a spiking borrow fee. The overlay can only reduce exposure, so no result here can be a leverage artifact.
How we kept the test honest
Halving positions mechanically lowers risk, which flatters Sharpe ratios and drawdowns whether or not the signal knows anything. So alongside the standard performance statistics, every rebalance records whether each position was flagged, then checks one holding period later whether flagged positions actually went on to do worse than unflagged ones. That conditional test is the one that can't be gamed by simply de-risking.
To make sure no conclusion hinged on one arbitrary choice, the whole experiment was swept across three rebalance cadences and three trend lookbacks: 36 backtests in total.
What came back
Three findings, in decreasing order of confidence.
The trend rule itself doesn't earn its keep here. It trails buy-and-hold SPY in every one of the nine sweep configurations, and its market-adjusted return is statistically zero. An earlier version of this study on mega-cap tech names looked far healthier; that strength came from the stocks themselves.
The overlays don't help. None adds significant return over the baseline. The best of them, the insider overlay, added less than half a percent a year, indistinguishable from zero.
One thread is worth remembering. Positions flagged for unusually sour news did go on to do slightly worse, and the effect strengthens at slower rebalancing. But it is small, it never converts into portfolio return, and it was found by looking across many cells. We log it as a hypothesis for a future pre-registered test rather than as a result.
Added return is each overlay's active return over the trend-only baseline; the p-value in parentheses is the probability of an effect this large by chance. Nothing approaches the 0.05 significance bar. For reference, SPY returned +236% over the same period.
Where this leaves us
The motivating question has a clean answer: no, these free alt-data streams do not improve this trend model. Just as importantly, the conditional scorekeeping shows why the overlays look good in naive statistics (halving positions flatters risk metrics regardless of skill) and prevents that from being mistaken for signal.
A study that ends in "no" still earns its cost. It closed a tempting avenue with measured confidence rather than lingering doubt, and it produced one disciplined follow-up question, the slow-cadence news effect, which we took to a broader test in Note 02.
Period 2019-01-01 to 2026-06-30. Panel: JPM, BAC, GS, CAT, BA, DIS, GE, UBER, QCOM, DAL. Data: Quiver insider transactions, Tiingo news sentiment, Interactive Brokers borrow fees, all point-in-time. Each metric is judged against roughly one year of the stock's own history, never cross-sectionally. Significance from weekly returns with Newey–West standard errors; conditional splits from a two-proportion test. Backtested results are hypothetical and gross of borrow costs; this note is research and is not investment advice.