KREN
← Yeongjun Yoo
WorldQuant BRAIN · Part 4 of 102026-06-29

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.

Multiplying signals, and the real rules of the scoreboard

A public record of building up "submittable alphas" to raise my score on the quant trading platform WorldQuant BRAIN. In part 3, I built my first submitted alpha out of mean reversion in price-to-cash-flow (pcf) and cleared the gate. This time, I hunted for a second alpha and checked firsthand how the scoreboard actually works. And I learned the real weight behind the rule that only one look-alike alpha gets through.

0. One-line summary

Instead of adding a weak signal orthogonal to the validated one (pcf), I multiplied them, and Sharpe jumped from 1.57 to 1.91. The intersection of stocks that are "cheap while the earnings outlook is being revised upward" became my second submitted alpha. And when I opened up the scoreboard, the rumor I'd heard, that "only one alpha counts toward your score per day," turned out to be wrong; the real rule was something else entirely. My first score was Bronze, 1,854 points.

1. Starting point: I passed the gate, but the score was nowhere to be seen

In part 3, I submitted pcf. It was an alpha that had cleared all 8 gates. But once I'd actually submitted it, I had no idea what was showing up on the scoreboard or how. Each alpha gets a grade, but it wasn't even clear whether that was the same thing as the competition score or something different. Without knowing the scoring rules, I couldn't judge "which alpha to submit when for maximum advantage." So this session had two goals: find a second alpha that didn't resemble pcf, and verify how the scoreboard actually worked using real data.

First, I went alpha hunting. At the end of part 3, I had a priority queue of 34 candidates built by 17 agents. I couldn't simulate all the strong-looking candidates at once. This account can run only one simulation at a time, so I had to line the queue up and go through it serially.

2. Working through the queue serially: alone, every signal was weak

I ran the queue's candidates one by one. Here are the results, by signal axis. The key point is what data I pulled, how, and what signal I turned it into.

Not a single candidate in the whole queue cleared the gates on its own. The only strong one was the option call-put implied-volatility spread I'd already met in part 3 (Sharpe 1.83), but that was a card I'd set aside because it was concentrated overfitting that collapsed in the sub-universe. Every orthogonal signal was weak on its own. This was the wall in this hunt. There were plenty of signals that didn't resemble pcf, but the ones that didn't resemble it were weak.

3. Breakthrough: don't add, multiply

I'd learned something in part 2. Adding a weak signal to a validated one at equal weight doesn't diversify, it dilutes. The strong signal got dragged down by the weak one. So this time, instead of adding, I decided to try multiplying.

Addition and multiplication mean different things. Addition means "bet if either one is strong," so the weaker side adds noise. Multiplication means "bet only when both point the same way," which creates an intersection. If you convert both signals into per-stock ranks and then multiply them, only stocks that rank high on both sides survive; a stock that's high on only one side gets suppressed.

The partner I chose to multiply by was the change in analysts' EPS estimates. pcf looks at "has the price-to-cash-flow ratio gotten cheap relative to its own history." I multiplied that by "are analysts revising the earnings outlook upward."

rank( -ts_zscore(cap / cashflow_op, 63) )
  × rank( ts_delta( ts_backfill(anl4_fs_detail_estimates_basic_af_v4_nd_eps_mean, 120), 66 ) )

anl4_fs_detail_estimates_basic_af_v4_nd_eps_mean is the analyst consensus EPS estimate. Since this data updates only sporadically on a quarterly basis, there are many empty days, so I filled in the last value with ts_backfill and used ts_delta to look at the change relative to 66 trading days (about one quarter) earlier. If this is positive, it means the outlook is being revised upward.

What this multiplication targets is a value trap filter. Sometimes a stock looks cheap because it's actually going under, and analysts are cutting their outlook on names like that. Conversely, a stock that "has gotten cheap, yet the outlook is actually being revised upward" is more likely to be genuinely undervalued. The intersection of cheap and improving. Multiplication pinpointed exactly that intersection.

The result was Sharpe 1.91, Fitness 1.6. Clearly higher than pcf alone (1.57). Adding had diluted it, but multiplying made it stronger. The strategy of multiplying an orthogonal signal into a validated base hit exactly as intended.

4. Clearing the second gate

I ran the server's 8 checks again. All of them passed. Turnover was low at 11.3% (since it's a fundamental signal, it naturally changes slowly), and the sub-universe Sharpe was also solid at 1.61. This wasn't an alpha that only happened to work in the broad market.

The most important gate was self-correlation. The second alpha couldn't be too similar to the first one (threshold under 0.7). Measuring how similar it actually was to pcf gave 0.61. Even though it used the same cash flow ingredient as its base, the multiplied-in analyst signal was different enough that the whole thing moved away from pcf. Under 0.7, so it passed. The second submitted alpha (cau_pcfleg) was locked in.

rank(-ts_zscore(cap / cashflow_op, 63)) × rank(ts_delta(ts_backfill(...eps_mean, 120), 66))
USA / TOP3000 / delay 1 / decay 4 / industry neutralization / weight cap 2%
→ Sharpe 1.91, Fitness 1.6, turnover 11.3%, sub-universe 1.61, self-correlation with pcf 0.61, 8/8 passed

5. Opening the scoreboard: the rumor was wrong

Now for the second goal: how the score is actually calculated. I'd vaguely assumed that an alpha's grade was the same as the score, but querying the score directly showed that wasn't the case. The competition scoreboard was a separate thing entirely.

Here's the rule set I confirmed.

This "average quality" thing is subtle. Submitting a weak alpha alongside a good day's batch drags down that day's average, a net loss. But cramming several strong alphas into a single day gets clipped by the 2,000-point cap. So the conclusion is: spread strong alphas across different days, and don't mix a weak alpha into a good day. I also confirmed that the score itself doesn't wait for future data (out-of-sample); it's applied immediately at the next update after submission.

The levers for boosting quality also became clear. The score rises the stronger the sub-universe result, the lower the self-correlation, and the higher the Fitness. Self-correlation in particular doesn't stop mattering once you clear the 0.7 pass threshold; the lower it is, the more it keeps working in your favor on the score. The upper pool (Power Pool) demands self-correlation of 0.5 or below outright. Collecting alphas that don't resemble each other wasn't just a matter of passing the gate, it was a matter of score too.

My first score posted. Bronze, 1,854 points. pcf alone had nearly filled the daily cap on its own. The second alpha's score was set to land at the next day's update.

6. Only one look-alike alpha: the trap of the cau family

Riding the momentum of the second alpha, I went after a third. Since multiplication had worked, I figured I could just make one more similar multiplication. I built a variant (cau_rev132) that multiplied EV/EBITDA mean reversion by the same analyst signal. Measuring self-correlation gave 0.42, which looked like a comfortable pass.

But there was a trap. That 0.42 had been measured before submitting the second alpha. After I formally submitted the second alpha and measured again, the third candidate's self-correlation had spiked to 0.83. A failure well past the pass threshold.

The reason was clear. The second and third alphas shared the same analyst revision signal (the revision leg) as their multiplication partner. Only the base (pcf or EV) differed; the multiplied side was the same, so the two were effectively cousins. The moment I submitted the second one, the third one that resembled it got blocked.

Here the bigger picture came into view. pcf, and the cau family formed by multiplying it with analyst revisions, were so similar to each other that alphas #1 and #2 had already saturated it. Making another one from this family runs into the self-correlation wall. The third alpha needs to be found outside this family, in a genuinely different signal that is neither price nor analyst revision. That became the homework for the next part.

7. What I learned this time

In the next part, I hunt for a third signal outside the pcf family, one that uses neither price nor analyst revisions. It's the hardest installment yet, discovering a new alpha that's dissimilar yet strong and robust.

Technical notes (appendix)