Factor Investing and Signal Fusion: Combining Alpha Signals

Direct answer: What is signal fusion in investing?

Factor investing and signal fusion combine many signals into a ranked conviction score while controlling overlap, decay, crowding, and portfolio fit.

How Celestice helps

Celestice connects research evidence, valuation assumptions, factor exposures, and portfolio fit. In practice, Celestice helps teams move from a market insight to a cited thesis, a scenario, and a portfolio decision with the caveats preserved.

One signal is noise; the edge is in the combination

Every investor has heard that some indicator "predicts" returns — a value factor, a momentum signal, an earnings-revision trend. Taken alone, almost any single signal is too weak and too noisy to trade on with confidence. The durable edge, the thing professionals call alpha, comes from combining many weak-but- real signals into one stronger conviction — and then proving that conviction holds up out of sample. That process is signal fusion, and understanding it separates disciplined quantitative investing from chasing the indicator of the month.

Factors and signals: the raw ingredients

It helps to separate the layers. Factors are broad, well-studied drivers of return — value, momentum, quality, size, low volatility — that explain why groups of securities behave the way they do. Signals are more specific, often faster-moving inputs: an earnings surprise, a change in analyst estimates, a shift in price trend. A strategy typically tracks a set of active factors and a set of active signal inputs, each contributing a piece of the picture.

The critical discipline is to keep these inputs honest: monitoring how many factors and signals are actually active, and watching for ones that are degrading — losing their predictive power as markets adapt and an edge gets arbitraged away.

Fusion: from many inputs to one conviction

Fusion is the step that combines those raw inputs into a single, ranked conviction per security. The naïve approach — equal-weighting everything — ignores that signals overlap and vary in reliability. A thoughtful fusion configuration weighs each input by its strength and, crucially, accounts for correlation between signals. Two signals that always agree are not two pieces of evidence; they are one piece counted twice. High signal correlation is a warning that apparent diversification of evidence is an illusion. Good fusion separates raw signal generation from the fused conviction it produces, so you can always see both the ingredients and the result.

Backtests: necessary, and dangerous

Before a fused strategy goes anywhere near real money, it has to be backtested — run against historical data to see how it would have performed. Backtesting is essential and also the single easiest thing to fool yourself with. The classic traps:

• Overfitting. Tune enough parameters and any strategy looks brilliant on the past while having learned nothing about the future.
• Look-ahead bias. Accidentally using information that would not have been available at decision time inflates results.
• Survivorship bias. Testing only on companies that still exist ignores the ones that failed, flattering the record.

A backtest is evidence, not proof. Its job is to disconfirm a bad strategy, not to crown a good one.

Backtest-to-live drift: the silent killer

Here is the metric that matters most and gets watched least: the drift between backtested and live performance. A strategy that looked excellent in simulation but underperforms once deployed is sending a clear signal — the backtest captured noise, the market has changed, or the edge has decayed. Monitoring backtest-to- live drift, alongside degrading factors and rising signal correlation, is how you catch a dying strategy before it does real damage rather than after.

Keeping the layers honest

The throughline of serious alpha work is separation of concerns: raw signal generation, fused conviction, backtest evidence, and live execution feedback are distinct layers, each inspected on its own terms. When performance disappoints, that separation lets you diagnose where it broke — a degraded factor, an over-correlated signal set, an overfit backtest, or genuine regime change — instead of throwing out the whole system blindly.

This is naturally suited to a continuous, agentic approach: track the active factors and signals, flag degradation and correlation as they emerge, surface pending backtests and approvals, and route strategy changes through review with the evidence attached — so conviction is always traceable back to the signals that produced it.

The takeaway

Alpha is not a single magic indicator; it is the disciplined fusion of many imperfect signals into one conviction, weighted for strength and correlation, validated by honest backtests, and continuously checked for the drift between simulation and reality. Respect each of those layers and you have a repeatable process. Skip any of them — especially the drift check — and you have a backtest that looks great right up until it loses money.