Every signal is wrong some of the time. The question is whether signals are wrong at the same time.

A single trading signal — even one with a validated edge — makes a probabilistic claim. A 60% win-rate signal is correct six times out of ten and incorrect four times out of ten. Over a thousand trades, those four-in-ten instances accumulate into a meaningful drag on performance. Some of those errors are recoverable: small losses, quick reversals, flat exits. Others are not: the 40% false positive that catches the trade at precisely the wrong moment in a high-volatility environment, or the one that holds a position through an earnings announcement the signal had no mechanism to anticipate.

The architecture question is whether you can reduce the false positive rate without proportionally reducing the win rate. Signal fusion — combining two or more genuinely uncorrelated signals — is one of the few approaches that does.

Why Single Signals Have High Noise

A single signal operates on a single information source. A moving average crossover reads price momentum. A relative strength indicator reads price velocity. A news sentiment score reads the information environment. Each reads one dimension of market reality and makes a prediction about future price behavior based on that dimension alone.

The limitation is not the signal's accuracy in isolation. It is the signal's inability to distinguish between its pattern occurring in a favorable context and its pattern occurring in an unfavorable context. A momentum signal fires whenever price momentum is above a threshold — regardless of whether that momentum is driven by sustained institutional buying or by a temporary volume spike in a thin market. A sentiment signal fires whenever the news is positive above a threshold — regardless of whether the broader market regime is risk-on or risk-off.

False positives are not random. They cluster around the conditions the signal was not designed to handle. Momentum signals produce false positives in choppy, non-trending markets. Sentiment signals produce false positives when a strong positive news event fires into a deteriorating broader market. Mean-reversion signals produce false positives when what looks like an overextension is actually a genuine structural break.

Over thousands of trades, these clustered false positives are the primary driver of the gap between backtested and live performance. The backtest period may have contained fewer regime mismatches. The live environment will contain more. Single-signal strategies have no mechanism for filtering these cases.

The Mathematics of Combining Uncorrelated Signals

Two signals, each 55% accurate, operating on independent information sources. The question is: what happens to accuracy when you require both to agree before entering a trade?

Under independence, the probability that both signals are wrong at the same time is 0.45 × 0.45 = 0.2025. The probability that at least one signal is correct is 1 - 0.2025 = 0.7975 — roughly 80%. This is the combined signal's accuracy when you enter only when both agree: approximately 80% of the time when both signals fire in the same direction, the direction is correct.

The tradeoff is trade frequency. Two independent 55% signals will agree on direction in roughly (0.55 × 0.55) + (0.45 × 0.45) ≈ 50% of opportunities, meaning you take fewer trades. The trades you do not take — the ones where only one signal fired — are a mix of wins and losses that you have chosen to avoid. Because false positives cluster in the cases where signals disagree, most of the avoided trades contain a disproportionate share of false positives.

The direction the math pushes in is clear: fewer, higher-conviction trades outperform more, lower-conviction trades when the friction costs (spreads, commissions, slippage) are non-trivial. Automated execution has real costs per trade. A strategy that fires half as often but wins 80% of the time instead of 55% produces better risk-adjusted outcomes in most realistic cost environments.

Three signals extend the mathematics further. Three independent 60%-accurate signals, all required to agree, produce a false positive rate of 0.40 × 0.40 × 0.40 = 6.4%. That is a dramatic reduction from the 40% false positive rate of any single signal. The trade frequency reduction is also dramatic — three independent 60% signals will agree only roughly 28% of the time — but the signal quality when they do agree is substantially higher.

What "Uncorrelated" Actually Means

The mathematics above requires independence. Two signals are independent if knowing the state of one tells you nothing about the state of the other. In practice, true independence between trading signals is rare and worth examining carefully before assuming it holds.

A price momentum signal and a volume-weighted momentum signal are not independent. Both read price behavior, one weighted differently. Their false positives cluster in the same conditions — choppy markets, thin liquidity, trend exhaustions. Combining them produces modest improvement over a single signal, not the compounding benefit that genuine independence provides.

A price momentum signal and a news sentiment signal are substantially more independent. Price momentum reads the recent price history. News sentiment reads the information environment that will affect price in the near future. Their false positive patterns are different: momentum fails in non-trending environments; sentiment fails in information ambiguity. When both fire, the cases where both are wrong simultaneously are rarer than the cases where either is wrong alone.

A price momentum signal, a news sentiment signal, and a market regime filter are close to genuinely independent. The regime filter reads the aggregate market information environment. The sentiment signal reads the individual ticker information environment. The momentum signal reads the price history of the specific security. These three operate on different information sources, at different scales, and their failure modes do not strongly overlap.

Testing correlation between signals is straightforward: compute the historical correlation between each signal's values over your backtest period. Correlation near zero indicates low linear dependency. But correlation is not the complete picture — signals can have low linear correlation and still share nonlinear dependencies or exposure to common factors. A more robust test is to examine whether their false positives cluster in the same market regimes. If both fail primarily in crisis environments, they are not independent for the purposes of fusion.

The Practical Challenge

Finding genuinely independent signals is harder than it looks. The most common mistake is constructing two signals that read different indicators but share the same underlying factor exposure.

Two equity momentum signals on correlated tickers — both in the same sector, both highly correlated to the same index — are effectively the same signal expressed twice. The correlation in their underlying securities ensures that they fire together and fail together. The fusion mathematics breaks down because the independence assumption fails.

Signals derived from different data sources can still share factor exposure. A sentiment signal for semiconductor companies and a momentum signal for semiconductor companies are both correlated to semiconductor sector performance. In a sector-wide sell-off, both signals fail together regardless of how independently they were constructed. Factor exposure is the hidden correlation source that practitioners most often underestimate.

The practical constraint is that genuinely independent signal pairs are limited. Technical + sentiment is one. Price momentum + macro regime is another. Intraday microstructure + fundamental value has been researched extensively in academic literature. Within each domain — all-technical, all-sentiment, all-macro — signal independence is harder to achieve.

Temporal independence is a variation worth considering. A signal computed on a one-day lookback and a signal computed on a twenty-day lookback are not fully independent, but their failure modes have less overlap than two signals operating on the same lookback period. Long-short timeframe fusion is a practical approach when genuinely cross-domain signals are not available.

The Overfitting Trap in Signal Combination

Combining too many signals introduces its own failure mode: overfitting the combination.

When you combine five or six signals using weights optimized on historical data, the optimization finds a combination that fits the backtest period precisely. Each additional signal adds free parameters to the optimization. Free parameters can improve in-sample fit without improving out-of-sample performance — and often actively harm it by fitting noise in the historical data.

The result is a fusion strategy that looks significantly better in backtest than a single-signal strategy but performs approximately the same in live trading, because the combination weights are not stable. The interaction effects between signals are regime-dependent. A weight that was optimal in the 2019-2022 period may be counterproductive in a different volatility and correlation environment.

The practical limit is two to three well-understood signals. Two signals require one combination decision: the AND condition. Three signals require three pairwise AND conditions plus a three-way condition. Beyond three, the optimization surface has enough degrees of freedom that overfitting becomes the dominant risk.

Well-understood is the operative word. The signals being combined should each have an explainable mechanism: why does this signal have an edge, in what market conditions does it work, and in what conditions does it fail? If you cannot answer these questions for each signal individually, you cannot know whether the combination makes structural sense or whether it is a numerical artifact of the backtest.

The Oyamori Approach

Oyamori treats signal fusion as a first-class architecture principle, not an optional enhancement. The platform is built around three independent signal layers: technical patterns (price and volume), sentiment (Newsvibe), and regime (aggregate sentiment market index).

Each layer reads a different dimension of market information. Technical reads what price has done. Sentiment reads what information is available about the underlying security. Regime reads whether the macro information environment supports strategy execution at all. These three layers have different failure modes, different data sources, and different temporal characteristics. Their combination produces the kind of structural independence that makes fusion effective.

Strategies built on the platform can express AND conditions across layers natively. A gap-and-go strategy can require: technical gap condition AND Newsvibe urgency flag AND regime is risk-on. All three conditions must be true for an entry to fire. The regime layer is always running — it cannot be disabled for a single strategy, because its purpose is to provide system-level protection regardless of what the individual strategy's signal is doing.

The sentiment and momentum fusion article covers the implementation detail of combining a specific sentiment signal with a specific technical signal — including threshold calibration, weighting decisions, and the practical look at how the combination performs in different regimes. That article is the applied companion to the architectural framework described here.

Algorithmic trading carries substantial risk. This article is educational, not investment advice.

Next: Sentiment + Momentum: Signal Fusion in Practice →