Seven Funerals and a Pulse

We've been hunting for an edge in the US stock market for about three weeks now. Here's the honest version of how that's going.

When we started this project, we were running sports prediction models — NBA spreads, soccer outcomes, the usual. The models worked okay. Not spectacularly, but okay. The problem was always the same: limited games, limited markets, limited opportunities to prove anything statistically meaningful. You'd wait an entire NBA season to get maybe 800 data points, and by then the market had shifted under your feet.

Stocks don't have that problem. There are thousands of tickers, filings drop every single day, and the data goes back years. More instruments, more signals, more chances to find — or disprove — an edge. That's what pulled us over.

So we started digging.

The Graveyard

First up: 8-K filings. The SEC publishes these every time something material happens to a public company — earnings, executive changes, asset sales, you name it. We scraped six months of data. 21,880 filings. Ran gap analysis on every single one.

The initial number looked incredible. PF = 2.39. We nearly celebrated.

Then we looked closer. The entire result was held up by two penny stocks — one that jumped 100% from a price of basically zero, and another that spiked 71% from $2. Remove the stocks you couldn't actually trade in real life (anything under $5 with no liquidity), and the profit factor dropped to **0.96**. Below breakeven. The edge was a mirage built on untradeable garbage.

Lesson: always check if your backtest results can survive contact with reality.

Next, we tried using LLMs to predict direction. We fed SEC filings to two different models — a large open-source one and a commercial one. Both scored **exactly 50% accuracy**. A literal coin flip. We ran 20 samples, then another 50. Same result. Turns out large language models are excellent at reading and extracting information, but they simply cannot predict which direction a stock will move. The entire academic literature agrees, and now so do we.

Then we found SHORT signals that looked genuinely promising. Three specific filing categories showed strong, statistically significant edges through walk-forward validation across 24 months of data. One had a profit factor of 3.04 in out-of-sample testing. We were ready to deploy.

One problem: you can't short stocks that aren't available to borrow. And it turned out the stocks generating our beautiful edge were precisely the ones nobody would lend you. Filter for actually-shortable stocks and the edge vanished completely. **PF went from 3.04 to noise.** Three months of work, dead in an afternoon.

Here's the full body count, because we believe in transparency:

| What We Tried | Result | Verdict |

|---|---|---|

| 8-K gap analysis (all stocks) | PF = 2.39 | Penny stock mirage |

| 8-K gap analysis (tradeable only) | PF = 0.96 | Below breakeven |

| LLM direction prediction | 50% accuracy | Coin flip |

| Rule engine (50 signals) | PF = 0.91 | Below breakeven |

| SHORT signal A | OOS PF = 1.16 | Collapsed out-of-sample |

| SHORT signal B | OOS PF = 3.04 | Can't borrow the stocks |

| SHORT signal C | OOS PF = 1.85 | Can't borrow the stocks |

Seven hypotheses. Seven funerals.

Something Survived

But the graveyard isn't the whole story. Somewhere in the middle of all this failure, we stumbled onto a LONG signal that passed every test we threw at it. Walk-forward validation, out-of-sample hold-out, bootstrap confidence intervals, robustness checks across every price band. The numbers didn't collapse when we removed outliers — they actually got stronger, which is the opposite of what happened with every failed strategy above.

We're not going to say what it is. Not yet. Maybe not ever, depending on how the next few weeks go.

What we will say is that it cleared our pre-defined Go/No-Go criteria by a wide margin, it survived seven skeptics arguing about it in a room for two hours, and it's now running in paper trading. Automated. Every night, the system scans for signals and places orders. Every morning, it closes positions. No human intervention required.

Four weeks of paper trading will tell us whether this is real or whether it joins the table above. We have hard kill conditions set in advance — if the numbers don't hold, we pull the plug automatically. No emotional attachment, no "let's give it one more week." The machine decides.

Why Stocks, Not Sports

Someone asked us why we pivoted from sports to stocks. The short answer: scale.

In sports, you might find an edge in NBA player props or soccer Asian handicaps. But you're limited to maybe 5-10 bets per day, seasonal schedules, and bookmaker limits that tighten the moment you start winning. The universe is small and it fights back.

In equities, thousands of events happen every single day. SEC filings alone generate 100+ actionable data points daily. If an edge exists, you can test it against tens of thousands of historical observations. And if it works, you can run it every market day without a bookmaker cutting your limits.

The flip side is that financial markets are brutally efficient. Every edge we thought we found in the first two weeks turned out to be either an artifact of bad data, a statistical ghost, or physically impossible to execute. But that's actually encouraging — it means when something does survive all the filters, it might be worth paying attention to.

We'll report back in four weeks with real numbers. If the paper trading fails, you'll see it in the graveyard table above with exact figures. We don't hide our losses.

---

*Not investment advice. Past performance does not guarantee future results. We may hold positions in securities discussed on this site. All analysis is for research and educational purposes only.*

This is not investment advice. Past results don't guarantee future performance. All analysis reflects our internal research process only.