Alpha Signals & Quant Research

What is alpha?

Alpha is a predictive edge beyond benchmark returns. A benchmark is the "default" return you get for simply holding the market (e.g. an index). Alpha is whatever excess return your strategy adds on top of that, because it actually predicts where prices are going.

Why might alpha exist?

If markets were perfectly efficient, there would be no edge to find. In practice, edges exist because markets are made of people and institutions, not perfect computers.

These frictions create temporary, exploitable structure in the data. A signal is just a rule that tries to capture that structure before it disappears.

Signal families

Most alpha signals fall into a handful of well-known families. They differ in what they assume about price behaviour.

flowchart TD
  A([Alpha Signals]) --> M([Momentum])
  A --> R([Mean Reversion])
  A --> B([Breakout])
  A --> V([Volume-based])
  A --> X([Cross-Sectional])
  M --> M1[Short MA > Long MA]
  R --> R1[Oversold bounce]
  B --> B1[Break of resistance]
  V --> V1[Volume spike]
  X --> X1[Rank assets: strong vs weak]
  classDef s fill:#eef2ff,stroke:#4f46e5,stroke-width:1px,color:#3730a3;
  classDef d fill:#ecfdf5,stroke:#0d9488,color:#0d9488;
  class A,M,R,B,V,X s
  class M1,R1,B1,V1,X1 d
      

Momentum signals

Momentum captures trend-following behaviour. The classic implementation is a moving-average crossover: compare a fast (short) moving average to a slow (long) one.

When the short MA crosses above the long MA, recent prices are outpacing the longer trend, so the rule goes long. This is the same logic behind breakout continuation — a strong move is assumed to persist.

flowchart LR
  P([Price history]) --> S([Short MA])
  P --> L([Long MA])
  S --> C{Short MA > Long MA?}
  L --> C
  C -->|Yes| BUY([BUY / Long])
  C -->|No| SELL([SELL / Flat])
  classDef s fill:#eef2ff,stroke:#4f46e5,stroke-width:1px,color:#3730a3;
  classDef d fill:#ecfdf5,stroke:#0d9488,color:#0d9488;
  class P,S,L,C s
  class BUY,SELL d
      

Mean reversion, volume & cross-sectional

Feature engineering

Signals are not built on raw prices directly. First you engineer features: transform raw market data (OHLCV, volume, macro data) into numeric inputs that may carry predictive information.

One handy feature is the volume ratio, a simple spike detector:

where $V_t$ is today's volume and $\operatorname{avg}(V)$ is the recent average (e.g. a 20-day average). Flag a spike when $\text{VR}$ exceeds a threshold, for example $\text{VR} > 2$.

Evaluating a signal

Before risking money, you measure whether a signal actually predicts anything. These are the core evaluation metrics — and the ones most likely to appear on the exam.

Hit rate

The fraction of trades that were profitable.

Average return per trade

Hit rate alone can mislead, so we also measure the average profitability across all trades.

Information coefficient (IC)

The IC measures how well a signal's predictions line up with the actual returns that followed — it is the correlation between the two.

An IC near $0$ means the signal is no better than random. Closer to $+1$ means strong predictive power; a negative IC means the signal is predicting the wrong direction. In practice even a small positive IC, applied across many assets, is valuable.

Stability, decay & costs

Volume spike detection

Generating a momentum signal

The quant research pipeline

Producing a signal is not a one-shot event — it is a disciplined, iterative pipeline. Most ideas die somewhere along the way, and that filtering is the research.

flowchart LR
  I([Idea]) --> H([Hypothesis])
  H --> D([Data])
  D --> F([Features])
  F --> S([Signals])
  S --> B([Backtest])
  B --> V([Validation])
  V -->|Robust| OK([Deploy])
  V -.->|Fails| It([Iterate])
  It -.-> H
  classDef s fill:#eef2ff,stroke:#4f46e5,stroke-width:1px,color:#3730a3;
  classDef d fill:#ecfdf5,stroke:#0d9488,color:#0d9488;
  class I,H,D,F,S,B,V,It s
  class OK d
      

The researcher's mindset

Key takeaways