What Are Quantitative Models?
Quant models find patterns. They don't understand markets.
Quantitative models do not forecast markets by understanding them. They identify statistical regularities in historical data and generate probabilistic assessments of whether those regularities are likely to persist into the near future. This distinction is not semantic. It determines what a quantitative model can legitimately claim, what it cannot, and how its outputs should be interpreted by anyone using them.
A quantitative model, in the context of financial markets, is a mathematical framework that takes structured numerical inputs, such as price history, volume, volatility measures, and macroeconomic indicators, applies a defined set of transformations, and generates an output: a signal, a classification, a probability estimate. The model does not know what a company does or why an earnings miss matters. It knows, statistically, what tends to happen to certain data patterns in certain conditions.
The intellectual lineage of quantitative modelling in finance is long. Harry Markowitz formalised portfolio optimisation using statistical inputs in 1952. Fischer Black and Myron Scholes developed options pricing models from stochastic calculus in 1973. William Sharpe's Capital Asset Pricing Model provided a factor framework for expected returns in the same decade. The firms that became legends of systematic trading, Renaissance Technologies, D.E. Shaw, Two Sigma, applied this logic at industrial scale beginning in the 1980s and 1990s. Modern machine learning-based models are the current iteration of a methodology that has been developing for seven decades.
Factor models, statistical arbitrage, and machine learning: three distinct approaches
Not all quantitative models are the same. Three broad approaches have shaped the field, and they operate on different principles.
Factor models attempt to explain the returns of an instrument through its exposure to a defined set of systematic risk factors. The Fama-French three-factor model, for example, extended the Capital Asset Pricing Model by adding size and value factors alongside market risk. More recent multi-factor models incorporate momentum, quality, low volatility, and dozens of additional factors. The signal generated by a factor model is a score reflecting the instrument's aggregate factor exposures, weighted by the historical predictive strength of each factor.
Statistical arbitrage approaches look for mean-reversion or co-movement patterns between instruments that history suggests tend to be statistically stable. When two historically correlated instruments diverge, the model generates a signal based on the expectation that the divergence will resolve. These approaches are highly sensitive to regime: what was statistically stable in one market environment may not be in another.
Machine learning-based signal generation, the most recent and increasingly dominant approach, trains models on large datasets of historical price, volume, fundamental, and alternative data to identify non-linear patterns that factor models would not capture. Rather than defining the factors in advance, the model learns which combinations of features are historically associated with subsequent directional movement.
How a quantitative signal is actually generated
The signal generation process involves several stages that are worth understanding concretely.
The first is feature engineering: transforming raw market data into the numerical inputs the model will use. A price series might be converted into a rolling volatility measure, a momentum score over various lookback windows, a ratio of recent volume to average volume, and a measure of where current price sits relative to its historical distribution. Each of these engineered features captures a different dimension of the instrument's behaviour.
The second is model training: exposing the model to historical data and adjusting its parameters to minimise prediction error across that dataset. This is where overfitting risk is most significant. A model that is too complex can learn the specific noise of the training data rather than the underlying signal, producing apparent accuracy in-sample that collapses out-of-sample.
The third is out-of-sample testing: evaluating the model's performance on data it was not trained on. This is the honest test of whether a model has learned something generalisable about market behaviour or merely memorised historical patterns. Rigorous quantitative research publishes and examines out-of-sample performance explicitly.
The fourth is signal calibration: converting the model's raw output into a probabilistic assessment with an associated confidence level. A well-calibrated model's 70% confidence signals should be directionally correct approximately 70% of the time, not more and not less. This calibration is what separates a usable signal from a raw model output.
The Trend Signal and what it actually represents
Opes Borsa's Trend Signal is the output of this pipeline applied to the instruments covered by the platform. It is a probabilistic directional assessment, positive or negative, accompanied by a Signal Confidence Score: the model's assessed probability of the stated direction based on current data.
The Signal Confidence Score is not a promise. A 92% confidence score means the model's historical calibration suggests that signals at this confidence level have been directionally correct at approximately that frequency on the data the model was developed and tested on. It does not mean the current signal will be correct. Markets can change in ways that shift the statistical regularities the model learned. The signal is an input to analysis, not a substitute for it.
The Emotionless Edge, expressed technically
The structural advantage of a quantitative model over a discretionary human analyst is not that the model is smarter. It is that the model is consistent. It applies the same feature engineering, the same trained parameters, and the same calibration framework at market close on the day of a major geopolitical event as it does on a quiet Tuesday. It does not widen its uncertainty estimates because the news frightened it. This consistency is the Emotionless Edge, and it is not a brand claim. It is a mathematical property of algorithmic systems.
Key Terms:
Quantitative Model: A mathematical framework that takes structured numerical inputs from financial markets and generates probabilistic output signals. It identifies statistical regularities in historical data rather than understanding the fundamental causes of market behaviour.
Factor Model: A quantitative approach that explains instrument returns through defined systematic risk factors such as market exposure, size, value, and momentum, weighted by their historical predictive strength.
Signal Confidence Score: In the Opes Borsa platform, the percentage figure accompanying each Trend Signal indicating the model's historically calibrated probability of the stated directional assessment being correct. Not a guarantee; a calibrated probability.
Overfitting: The risk in model training that a model learns the specific noise of its training dataset rather than generalisable patterns, producing apparent in-sample accuracy that collapses when tested on new data.
Out-of-Sample Testing: The evaluation of a quantitative model's performance on data it was not trained on. The standard of intellectual honesty in quantitative research, distinguishing models that have learned genuine signal from those that have memorised historical patterns.
