The most counterintuitive thing I learned building quant systems:
Being right about direction is not the same as making money.
I've watched traders correctly call the outcome of 6 out of 10 trades and still end the month negative.
Not because of bad luck. Because of a fundamental mathematical error in how they were building conviction.
This article is about that error - and the institutional framework that eliminates it.
THE PROBLEM NOBODY EXPLAINS
Every prediction market trader has experienced this:
You analyze the situation.
The evidence points clearly in one direction.
You enter with confidence.
The market moves against you.
You were right in the end - but too early, or sized too large, or the timing was off.
You blame execution. You blame the market. You never blame the model.
But the model was the problem.
Every signal has an Information
Coefficient - the correlation between what it predicts and what the market actually does.
The best individual signals used at institutional desks sit between 0.05 and 0.15.
Wrong the vast majority of the time. Not occasionally. Structurally.
And yet those desks print money consistently.
The reason is one equation.
THE FUNDAMENTAL LAW
IR = Information Ratio of your combined system - your risk-adjusted edge
IC = average Information Coefficient of individual signals
N = number of genuinely independent signals
If each of your 50 signals has IC 0.05:
Single signal with IC 0.10 running alone:
The 50-signal system at half the individual strength is 3x more powerful.
This is why hedge funds employ hundreds of researchers to build hundreds of signals.
The one perfect indicator doesn't exist.
The system of many imperfect ones does.
WHAT COUNTS AS A SIGNAL
Five categories:
1. Momentum signals
2. Mean reversion signals - deviation from expected fair value relative to comparable instruments
3. Volatility signals - implied vs realized volatility gap
4. Factor signals
5. Microstructure signals
When spread expands: someone is trading on information the market hasn't priced yet.
THE 11-STEP COMBINATION ENGINE
Step 1 - Collect historical return series R(i,s) for each signal i and period s
Step 2 - Serial demeaning:
Step 3 - Sample variance:
Step 4 - Normalize:
Step 5 - Drop most recent observation
Step 6 - Cross-sectional demeaning:
Step 7 - Drop one more period
Step 8 - Expected forward return:
Step 9 - Regress E_normalized on Lambda. Take residuals e(i). These represent each signal's genuinely independent contribution.
Step 10 - Set weight:
High independent edge and low noise gets more weight. No subjective judgment.
Step 11 - Normalize so sum of absolute weights equals 1.
WHY CORRELATION IS THE REAL ENEMY
The N in IR = IC x sqrt(N) is not the count of signals.
It is the effective number of independent signals after shared variance is removed.
Running 50 correlated signals gives you the diversification benefit of 10-15 independent ones.
A portfolio manager who believes she's running 20 independent signals may be running 6.
Position sizes justified by 20 independent views are far too large for 6.
That leverage mismatch is the mechanism behind most systematic strategy blowups.
APPLYING THIS TO POLYMARKET
polymarket.com/?r=atlas - $28 billion traded, 9,000+ active markets.
Five probability signals for Polymarket:
1. Cross-venue pricing - spread between Polymarket and Betfair
2. Calibration signal - historical resolution rates across 400M trades
3. Bayesian update:
4. Microstructure - informed order flow via VPIN
5. Momentum - rate and direction of price movement near resolution
Before automating - build intuition for correlated markets manually.
PolyParlay is the first Telegram bot for Polymarket parlays. Combine any markets into one position:
t.me/poly_parlay_bot?start=ATLAS
Every correlation you notice in a parlay becomes a potential independent signal worth testing.
POSITION SIZING
Empirical Kelly adjusted for uncertainty:
Run 10,000 Monte Carlo path simulations. Measure how much your edge estimate varies. The more it varies - reduce your Kelly fraction.
SUMMARY
Being right about a market is not the same as having a profitable model.
IR = IC x sqrt(N)
The 11-step combination engine gives you optimal weights that reflect each signal's independent contribution, penalize noise, and remove shared variance.
The traders who consistently lose on trades they were right about are almost always losing to correlation they didn't measure.
The combination engine removes that failure mode structurally.
Now you know the engine.
