An “8 hours, outside the box” directive. We crossed fundamental states with technical states across three
universes (large caps, mid/small caps, and ~4,500 listed names outside the big indexes) and ran each cell
against its own random baseline.
First, the humbling control
We ran a mass search of 614 technical-indicator combinations per index — and ran the same search on
20 statistically faked price histories. The best “discovery” on real data improved Sharpe by +0.35; the
median best on fake data was +0.38. In other words, mass technical-combination searching finds the same
quality of “edges” in noise as in reality. That control is now standing policy: any mass search needs a
fake-world baseline before anything it finds is believed.
What actually validated
- Strong fundamentals plus an early-uptrend technical state, in the universe outside the big
indexes: on the order of +17–27%/yr over that pool’s baseline with solid statistics. Technicals time the
fundamental edge; they don’t create one (the same technicals on all names: nothing). - Operating-inflection signatures (“about to turn profitable, revenue accelerating”) on liquid non-index
names: ~+19–25%/yr at 6-month holds across several specification variants — a plateau, not a spike. - Insider cluster-buying inside the S&P 1500: ~+11–14%/yr versus baseline. The same signal in
microcaps is negative — insider buying in tiny names reads as distress/promotion, not information.
Why watchlist-only
These cells live in thinner names where our tradeability gates eat much of the edge, and survivorship bias
is loudest exactly there. So the signatures run as an automated Saturday scan that feeds a public-side
watchlist and the conviction-idea pipeline — candidates for hand research, never automatic buys.