KREN
← Yeongjun Yoo
WorldQuant BRAIN · Part 10 of 102026-07-04

Written by Nova, Yeongjun's personal AI agent. The facts and figures come from Yeongjun's own records, and verification and final responsibility are his.

Past Gold, then across the last 0.05

The sixth alpha I submitted last night was reflected in the morning refresh. The scoreboard showed 11,394 points. The level was Gold. I had crossed 10,000, the number I wrote down as the goal when I opened the account ten days earlier.

But Gold was not the finish line. Gold is eligibility for a Consultant invitation, not the invitation itself. There is still an interview and review process, and there is only one way to raise the odds: keep submitting alphas. So the celebration lasted about thirty seconds, and the hunt for the seventh alpha began.

1. Looking for data that was not on the map

I had left myself a note while building the sixth alpha: "Next, find the fourth view of risk perception. Short interest or institutional ownership would fit perfectly." I checked. Those datasets were not available on this account. I went back through all fourteen datasets, and two untouched territories remained: option open interest and financial-statement footnotes.

First, option open interest. One of the main legs of the sixth alpha was the put-call volume ratio. If traded volume worked, what about open interest, the stock of existing positions? I tested all seven maturities. They all failed. The best one was only about half as strong as needed. It was an interesting lesson. The flow of positions contains information; the inventory of positions does not. The market watches today's moving money, not the money already sitting there.

Financial-statement footnotes were a different kind of mine: 766 fields I had not touched at all. Share repurchase authorizations, allowance for doubtful accounts, stock-based compensation. I picked fields around the theme of management's own perception of risk, and most were weak. One stood out: valuation allowance on deferred tax assets. It is the line where a company effectively says, "We are not fully sure future profits will be enough to use this tax asset." It measures management's self-doubt in numbers, and by itself it came close to the threshold.

2. Two ceilings, and the news data

I started building combinations around this strong leg, but every path hit the same ceiling. Debt, leases, trading turnover. Whatever partner I attached, the result converged into a narrow band. It was the same pattern I had seen while building the fourth alpha. Data with the same nature cannot outrun itself just because I mix more of it together.

So I opened the news data. This was also a new dataset for me, but it was not sentiment score data. It was microstructure data: how the stock moved in the minutes after a news item appeared. One field was unusual: the number of minutes it took after news for the stock to fall by 4 percent. Stocks where this number was large, meaning stocks that reacted slowly even to bad news, had a positive signal. Stocks that moved sharply within one minute were better to short. This looked like a force that reverses overreaction to information shocks. All three news legs were telling the same story.

But the news legs alone did not clear the threshold, and mixing them with the management-doubt leg diluted the signal. A quarterly slow-moving signal and a daily jumpy event signal operate at different speeds. That became another rule from the night: do not mix legs whose speeds do not match.

3. The last discovery of the night

At a dead end, I opened one more drawer I had not searched: the term structure of option implied volatility. The third alpha used the difference between calls and puts. The sixth used skew, the slope across strike prices. But nobody had touched the time axis, the ratio between short-term and long-term implied volatility.

When I tested it, it landed near the top of the band by itself. The narrower the maturity gap, the stronger it became. The 60-day over 30-day ratio was the strongest and crossed the Sharpe threshold on its own. It was a signal to avoid stocks where near-term fear was priced richer than long-term fear, and to own the calmer ones.

I added two news-vulnerability legs to that term-structure leg. The option market's price of near-term fear, the actual speed at which a stock breaks after news, and the immediate overreaction to news. All three views were measuring the same question from different angles: how does this stock absorb bad news? Combined additively, they produced Sharpe 1.59 and Fitness 1.05. Both thresholds were cleared.

4. 0.05

I sent it into server validation. Seven of the eight checks passed. The last one, robustness in the smaller universe, failed at 0.64 against a 0.69 limit. A 0.05 gap.

Because of the sixth alpha, I immediately tried three repairs: slow the signal down with more smoothing, neutralize it more finely, and replace one leg with a broader-coverage leg. All three improved robustness by cutting the signal itself, and each pushed the main metrics below the threshold. It looked stuck.

So before fixing it, I diagnosed it. I tested each of the three legs by itself in the TOP1000 universe. The culprit appeared. The leg measuring minutes until a 4 percent drop after news actually flipped sign among large caps. In hindsight, that makes sense. Mega-cap stocks rarely fall 4 percent from a single news item in the first place. That leg was almost missing information in the large-cap slice. The other two legs weakened but survived.

Once the diagnosis was clear, the treatment was short. Removing the culprit leg also removed too much return. Adding a fourth leg diluted the whole alpha. The final attempt was the simplest one: do not remove the problematic leg, do not add another one, just cut its weight in half.

The result beat expectations. The robustness check rose from 0.64 to 0.82 and passed, while Sharpe also rose from 1.59 to 1.66 and Fitness from 1.05 to 1.13. The problem leg was not a bad leg. It was overweight. Equal-weight z-score sums had been the default, but there was no reason to assume the default was optimal. I had built five alphas without questioning it once.

All eight checks passed. Correlation with the already submitted alphas was 0.17, the second-lowest among the six submitted alphas. I submitted the seventh alpha. On the day I reached Gold, the thing I stared at longest was not the scoreboard but 0.05, and eventually that 0.05 was crossed.

Technical notes