The thesis
News-sentiment runs a simple, legible rule across a 24-name large-cap universe (AAPL, MSFT, NVDA, JPM, JNJ, XOM and peers): buy when recent news sentiment turns positive, exit when it turns negative. The appeal is interpretability — every position should map to a headline-driven signal rather than an opaque factor stack. Whether the signal actually fires often enough to matter is the open question the data keeps returning to.
Recent activity
The strategy is live and trades on a scheduled evening run. Over the last week most sessions did nothing tradable: the runs on 24, 25, 28 and 29 September each executed zero orders, with one or two rejections apiece. The exception was 23 September, when it executed two orders — buying 4 shares of NVDA at $225.34 and selling its 35-share BAC position at $55.96, a lot it had opened back on 12 June at $55.93. Earlier executed buys this summer included GOOGL, UNH, MSFT and AAPL. The paper book currently sits at roughly $10,172 total with $2,036.71 in cash. Worth flagging: rejected orders recur in nearly every session, a sign the signal is generating intent the account can't fill — typically a cash or sizing constraint.
Backtest and validation
Here the picture turns sober. Across a 451-day backtest the strategy returned just 0.19% (final equity $10,018.55, ~0.1% CAGR) on only 2 trades, with a 0.05% max drawdown, a 0.74 Sharpe and a reported 0% win rate. Turnover was 36.5% and fees a trivial $2. Walk-forward validation did not pass: of four folds, three placed no trades at all, and only the final fold (Dec 2025–May 2026) was active, contributing the entire 0.19% return and a 1.5 fold Sharpe. So just one fold in four was positive. The probabilistic Sharpe (0.923) reads well, but the deflated Sharpe falls to 0.551 once you account for 6 trials — and on a two-trade sample, none of these figures carries real statistical weight.
Strengths and risks
The strengths are genuine but modest: the rule is transparent, realised drawdown has been negligible, and trading costs are minimal. The risks dominate. The core problem is inactivity — three of four validation folds never traded, so we have almost no evidence the signal triggers reliably across regimes. A two-trade, 0%-win-rate backtest is not a track record; it is noise. The recurring live rejections compound the concern, hinting the book struggles to act on its own signals.
The honest read: news-sentiment is worth keeping live to accumulate data, but it has not earned a larger allocation. Until it demonstrates it can consistently trade its thesis — and produce a sample big enough to measure — treat every headline metric here as provisional.