What is IC?
IC (Information Coefficient) is the core performance metric in SignalNet. It measures how well your predictions correlate with actual stock returns.
The Simple Version
IC answers one question: "When you said a stock would go up, did it actually go up?"
- IC of +1.0 = perfect prediction (never happens)
- IC of 0.0 = random guessing (no skill)
- IC of -1.0 = perfectly wrong (also never happens)
- IC of +0.02 to +0.05 = a very good quant signal
In practice, consistent IC above 0.02 across many rounds is exceptional. Most professional quant funds operate with IC in the 0.01–0.05 range.
How It's Calculated
IC is the Spearman rank correlation between your predicted rankings and the actual stock returns over the resolution period.
Step by Step
- You submit predictions: each stock gets a score from 0.0 (bearish) to 1.0 (bullish)
- After 20 trading days, we observe actual returns for each stock
- Both your predictions and actual returns are rank-ordered
- IC = Spearman correlation between the two rank orderings
Your predictions: AAPL=0.9 MSFT=0.7 GOOG=0.5 META=0.3 AMZN=0.1
Actual returns: AAPL=+8% GOOG=+5% MSFT=+3% AMZN=+1% META=-2%
Your ranks: AAPL=1 MSFT=2 GOOG=3 META=4 AMZN=5
Actual return ranks: AAPL=1 GOOG=2 MSFT=3 AMZN=4 META=5
Spearman correlation → IC ≈ +0.70 (strong agreement)
Why Spearman (Ranks) Instead of Pearson (Raw Values)?
- Scale-invariant — It doesn't matter if you predict 0.9 vs 0.99 for the top stock, only the ranking matters
- Outlier-resistant — One stock going up 50% doesn't dominate the score
- Fair comparison — Everyone is compared on the same basis regardless of prediction scale
The Three Scoring Metrics
SignalNet scores your submissions on three IC-based metrics:
IC (50% weight)
Correlation — the standard Spearman rank IC between your predictions and actual returns.
This is the most straightforward: did you predict the relative ordering of stocks correctly?
IC = spearman_correlation(your_ranks, actual_return_ranks)
A contributor with IC of +0.03 is doing well. IC of +0.05 is excellent.
TC (30% weight)
True Contribution — measures the unique signal you provide that the meta-model doesn't already capture.
If 100 contributors all predict the same thing, the meta-model already knows it. TC rewards predictions that are different from the crowd while still being accurate.
residual = your_predictions - meta_model_predictions
TC = spearman_correlation(residual, actual_returns)
High TC means: "You know something the crowd doesn't."
MMC (20% weight)
Meta-Model Contribution — measures how much the overall ensemble improves when your signal is included.
Even if your individual IC is modest, your signal might be highly valuable if it captures a dimension that other contributors miss. MMC measures this.
meta_model_without_you = average(all_other_predictions)
meta_model_with_you = average(all_predictions)
MMC ≈ IC(meta_with_you) - IC(meta_without_you)
Positive MMC means: "The ensemble is better because of your signal."
Final Score
Final Score = 0.5 × IC + 0.3 × TC + 0.2 × MMC
These weights can vary per round (the Genesis Round uses 70/20/10 to be more beginner-friendly).
What Makes a Good IC?
| IC Range | Interpretation |
|---|---|
| +0.05 or above | Exceptional — top tier contributor |
| +0.02 to +0.05 | Strong — consistently profitable |
| +0.01 to +0.02 | Decent — positive edge, compounding over time |
| 0.00 to +0.01 | Marginal — barely above noise |
| Below 0.00 | Negative — your model is systematically wrong |
Context from the Real World
- Renaissance Technologies (the most successful quant fund ever) is rumored to operate with IC around 0.05
- Most published academic factors have IC of 0.01–0.03
- A random model has IC ≈ 0.00 (by definition)
An IC of 0.03 doesn't sound like much, but applied across 503 stocks over hundreds of rounds with staking, it compounds into significant returns.
IC vs. Sharpe Ratio
In traditional finance, performance is measured by the Sharpe ratio (return per unit of risk). IC and Sharpe are related:
Sharpe ≈ IC × √(breadth)
Where breadth = number of independent bets (stocks × rounds).
With 503 stocks and weekly rounds, even a modest IC of 0.02 translates to a high strategy-level Sharpe ratio. This is the Fundamental Law of Active Management — many small edges compound.
Cumulative IC
On your dashboard, Cumulative IC is the running sum of your per-round IC scores. It shows your total predictive edge over time.
A rising cumulative IC line means your model is consistently finding signal. A flat line means you're breaking even. A declining line means your model needs work.
Tips for Improving Your IC
- Feature engineering matters more than model complexity — A well-crafted feature set with a simple model beats a complex model with bad features
- Neutralize sector exposure — If your model just bets "tech goes up," you'll have high variance and low IC in non-tech environments
- Ensemble your own models — Combining 3-5 diverse models often beats any single model
- Watch for overfitting — High IC on training data and zero IC on live rounds means your model memorized noise
- Focus on TC — Once your IC is decent, maximize TC by making your signal orthogonal to the meta-model