Skip to main content

Signal Quality & IC Tracking

TRADEOS.tech does not assume that any signal which worked historically will continue to work. The system continuously measures signal quality in live conditions and adjusts its behavior accordingly. This is one of the most important engineering decisions in a live systematic trading system — and one of the most commonly skipped.

The Information Coefficient (IC)​

The Information Coefficient is the Pearson correlation between a signal's predicted score and its actual realized outcome (PnL denominated in R — multiples of risk). An IC of +1.0 means the signal perfectly predicted outcomes; 0.0 means it was no better than random; negative means it was actively wrong.

IC is the single most important signal quality metric because it measures the only thing that matters: does the signal actually predict what it claims to predict?

How TRADEOS.tech Tracks IC​

For every signal type, TRADEOS.tech maintains a rolling buffer of (composite_score, realized_pnl_r) pairs — one pair recorded after each trade closes. The Pearson correlation across this buffer is the live IC.

  • The buffer has a fixed maximum depth so recent behavior is more influential than distant history
  • A minimum sample count must be reached before IC contributes (prevents noise from tiny samples)
  • IC is cached and refreshed periodically, not recomputed on every signal event

IC → Sizing Multiplier​

IC is translated into a multiplier applied to position sizing:

  • High IC: signal is performing well → multiplier elevated above 1.0
  • IC near zero: signal is unpredictive → multiplier near neutral
  • Negative IC: signal is actively harmful → multiplier at floor (significant reduction)

This means a signal that was previously reliable but has started decaying will automatically see smaller position sizes — without requiring manual intervention.

Calibration Drift Tracking​

A separate mechanism tracks calibration drift: whether a signal's stated confidence aligns with its actual win rate.

After each trade closes, the system records predicted confidence against actual outcome per signal type. A running time-weighted average of the gap between predicted and realized outcomes tracks drift:

  • Positive drift: signal is underconfident — it wins more than it predicts. Confidence adjustments are nudged upward.
  • Negative drift: signal is overconfident — it wins less than it predicts. Confidence adjustments are nudged downward.
  • Near-zero drift: signal is well-calibrated. No adjustment applied.

Drift alerts are published when the magnitude exceeds a threshold, allowing downstream systems to respond immediately rather than waiting for the next scoring cycle.

Drift state persists across restarts via Redis, with staleness decay applied to signals that have gone silent for an extended period.

Hero Signals​

Not all signals are treated equally. TRADEOS.tech identifies hero signals — a dynamic set of signal types that have demonstrated exceptional live performance across multiple quality dimensions:

  • Sustained high win rate over a meaningful sample size
  • Positive PnL contribution (net profitable, not just technically correct)
  • Strong calibration (stated confidence close to realized win rate)

Hero signals receive preferential treatment in the composite scoring layer: they carry higher weight in the confluence gate and may qualify for elevated sizing tiers.

Critically, hero signal status is dynamic. A signal can enter or exit hero status based on live performance. The system does not rely on static designations made during backtesting — it measures continuously and updates the set as evidence accumulates.

Confluence Velocity​

In addition to IC and calibration, TRADEOS.tech tracks signal arrival velocity per signal type: a time-weighted measure of how fast signals of a given type are arriving relative to their historical baseline.

When a signal type is arriving faster than its historical baseline — indicating a concentrated burst of evidence — the confluence velocity factor is boosted, increasing the effective weight of those signals in the confluence gate. When arrivals are sparse, the velocity factor is reduced.

This captures the intuition that a sudden clustering of signals from a particular alpha source is more meaningful than the same number of signals spread thinly over a long period.

What This Means in Practice​

These mechanisms work together to create a self-regulating system:

  1. A new signal enters with neutral IC and calibration (no history yet → conservative weighting)
  2. As trades accumulate, IC and calibration are measured continuously
  3. High performers receive increasing weight and sizing authority
  4. Underperformers are automatically down-weighted or excluded
  5. Hero signal status formalizes the best performers into the core ensemble

The system degrades gracefully: when data is sparse or a signal is new, it defaults to conservative treatment. As evidence accumulates, trust is extended — or withheld — based on observed performance, not assumptions.