The procedure for making discoveries within the field of quantitative finance is called quant research.

A quant strategy is a systematic strategy, driven by data and the model, which is programmed to generate investment decisions.

The main pillar within quant research is the scientific method.

What is the scientific method?

The scientific method is a method that allows us to categorize the relationships of phenomena in an objective way, and replicate them on the observed phenomenon.

The scientific method is based on the formulation of a hypothesis and an empirical evolution of the thesis, using high standards of analytical rigor, and using sensitivity analysis to "manipulate" the hypotheses and the context in which the premise was formulated, to obtain greater rigor in the conclusions.

Correlation does not imply causation.

Which means that a supposed association between variables does not necessarily directly cause a reaction.

Causality is an extra-statistical concept. It is given by the nature of the phenomenon, and not by mathematical events.

Phases of the scientific method within quant research.

• Phase 1: This phase, called the "phenomenological step," is the basis of observation. The search for recurrent patterns, which can initially be associated with events or phenomena, without going further. The main objective is to find an initial situation for our "problem." For example, we want to investigate whether a rise in the price of asset A affects asset B. The traditional approach has been through logical induction, but currently a wide variety of methods are used.
• Phase 2: At this point, we formulate hypotheses based on the observations made. An observable causal mechanism is proposed as responsible for the phenomenon.
• Phase 3: In this phase, called falsification, the objective is to falsify each component that has been categorized as part of the causal mechanism. It deduces the key implications of the theory, for example, it is impossible for the hypothesis to be true and the implications to be false. The main objective of this phase 3 of the procedure is to validate that the assumed implications are incorrect, through logical deduction, among many other methods. People with experience and creativity are needed in this phase.

Most Used Experiments

There are three types of experiments used.

import pandas as pd
import numpy as np

# Quant research pipeline: data -> features -> backtest -> deploy

# 1. Load market data
data = pd.read_csv("spy_daily.csv", index_col=0, parse_dates=True)

# 2. Generate features (alpha signals)
data["returns"] = np.log(data["close"] / data["close"].shift(1))
data["vol_20"] = data["returns"].rolling(20).std()
data["momentum"] = data["close"] / data["close"].shift(20) - 1
data["rsi"] = compute_rsi(data["close"], window=14)

# 3. Define trading signal
data["signal"] = np.where(data["momentum"] > 0.02, 1, 0)

# 4. Backtest: compute strategy returns
data["strategy"] = data["signal"].shift(1) * data["returns"]
sharpe = data["strategy"].mean() / data["strategy"].std() * np.sqrt(252)
print(f"Sharpe ratio: {sharpe:.2f}")

• Intervention Experiments: In a controlled environment, the effect of Y on the value of X is observed while keeping the rest of the variables constant. In addition, techniques such as RCT (Random Controlled Trial) are used to validate the tests.
• Natural Experiments: Classification into control groups determined randomly. Or over broad universes of assets. The objective is the most robust validation possible. Methods such as Cross-Sectional Studies (COS) or Difference-in-Differences (DID) are used.
• Simulated Interventions: Any methodology that uses synthetic series

Quant Factors or Quant Models

All models are wrong, some models are useful

Models are idealizations of the real world, attempting to reduce the complexity of the day-to-day to a few mathematizable components that can capture Alpha in some way. Models aim to describe the qualities from which profitability is intended to be extracted.

It is important to note that not all market dynamics are modelable, and that a model does not include all known mechanisms, but rather tries to isolate a specific component.

Within all models, there is no homogeneity; they differ in their motivation to trade, operating frequency, information used to trade, markets operated, holding periods, etc.

• Multifactor Strategies: They invest in assets based on multiple factors. Their main objective is to anticipate market decisions. These strategies derive their Alpha from their computing capacity, and with the new Nvidia GPUs, processing levels have reached levels unattainable for the retail user.
• Asset Allocation Strategies: Strategies with decisions based on allocation (total % of assets under management) in sectors, countries, factors, market regimes, timings, etc., and their advantage lies in their ability to switch between risk-on and risk-off scenarios.
• Stock-Specific Strategies. Strategies based on information from a single asset, and targeted exclusively at that asset.
• Single Factor Strategies. Strategy based on rules or conditions, using techniques to obtain Alpha from assets, usually via premium.
• Event Management Strategies: Any strategy that generates Alpha on anticipated events, such as US presidential elections, wars, or any unanticipated event that creates a market dislocation and enables the possibility of extracting Alpha.
• Market Microstructure Strategies. Value extraction from order flow opportunities and order book dynamics. Usually used in the very short term.
• Statistical Arbitrage Strategies. Systematic exploitation of relationships between assets with similar characteristics. Unlike pure "risk-free" arbitrage, statistical arbitrage does involve risk.
• Textual Strategies. Strategies that base their decision logic on LLM text models with information from news, reports, internal documents, internet searches, and other information sources.
• Thematic/Macro Strategies Asset basket strategies based on macro decisions on economic development.

There are two major groups of models

• Hypothesis-based, or deductive: The starting point is research into where a profitability opportunity exists, and why it exists. It depends on economic theses or internal market mechanics. Usually, it is preceded by empirical work.
  
• Inductive pattern-based: This approach has an exploratory nature, where it discovers emerging advantages from empirical work. One must be extremely critical and careful to distinguish between correlation and causation.
  The most classic models are:

How are quant strategies developed?

• Research and Formulation
• Signal Development
• Acquire and process the data
• Analyze the data
• Develop the strategy to go to market and leave it in test mode
• Analyze, test, and evaluate the quant model
• Implement live if everything is ok

Formulating hypotheses about trading ideas and strategies

To succeed in quant trading, it can be for two reasons: luck, or doing good work.

To do it through work, everything starts with an idea based on an economic intuition, an internal market behavior, or an anomaly.

Having access to a collection of research, both public and private, makes things easier, since public internet papers are like abandoned appliances in the trash—maybe with some repairs it could be functional, or maybe it's useless, but it's clear that someone has already used the anomaly until exhausting it, and subsequently published it.

Something logical considering the amount of resources and time devoted to preaching in the desert (the research work), until some Alpha is found somewhere.

Furthermore, differentiate between trading ideas and quant strategies. A quant strategy is oriented toward the long term, while a trading idea has a much shorter specific horizon, seeking a specific dislocation in an event, or mispricing, while quant strategies seek to exploit opportunities more associated with anomalies or dislocations.

Generating Signals

Signals create a framework for generating an investment strategy. Using the data we have formulated with, we generate a method to quantitatively measure the successes that represent the investment idea.

Although it varies depending on the idea or investment thesis, at this stage, within the Python quant ecosystem with zipline, the pyfolio library is typically used to perform this signal analysis.

Example of a signal: A positive produced by a news sentiment analysis. Signals do not imply immediate entry, but rather a state change—to look for an entry, or to follow the next instruction from the model.

Acquiring and Processing Data

The heart of every model, of all research, comes from the data. Data is the most critical part of the process. And with errors in the data, there are only errors in the studies and procedures.

• Big Data Management I: Introduction to big data for trading. [PDF] | @randominvesting
• Big Data Management II: First lines of code. [PDF] | @randominvesting
• Big Data Management III: Building a datalake on NASDAQ with ArcticDB [PDF] | @randominvesting
• Big Data Management IV: Using zipline with the datalake (arcticdb) [PDF] | @randominvesting

What is good data?

It is data that meets a series of mandatory requirements.

• That they are consistent in their definition and what they represent
• That they have sufficient detail about the data's underlying
• That they have adequate temporal availability
• That they are consistent over time
• That they are free of "look-ahead" bias or crystal balls
• That they are free of survivorship bias

Furthermore, when working with data and its storage, at a reputable asset manager, all its systems were required to work as follows:

• History of datalake development
• Understand standard datalake usage procedures, and the differences between different sources
• Risk report on data risks.
• Feed redundancy. Multiple feeds for the same asset, also very important for the backtesting stage, having information from different sources.
• Use of metadata, versions, and snapshots
• Outlier report and their economic explanation hypotheses.

The technical conditions are covered by the use of arcticdb, and the tacit ones do not require extremely special resources to carry them out successfully. Therefore, we can create a near-professional environment with open-source solutions and very low cost.

Signal Analysis

Within signal analysis, we focus on statistical tests and econometric techniques on the data and model results, to empirically evaluate the properties of the signals. Pyfolio is the ideal candidate for this task in most cases. It is the minimum Alpha unit needed to create more complete quant systems.

Building the Strategy

A model represents the mathematical or systematic rules of a trading system. In this case, we will move our model to a less abstract form. We will specify the model. We will work on signal selection and the combination of signals with each other, as well as exogenous conditions, such as economic inertia in the trading strategy (if the market is very choppy, stay calm... and things like that).

Evaluation, Testing, and Implementation.

In this final stage, we will estimate the quality of the model and the quality of the forecasts. Our analysis will include how well it fits in-sample and out-of-sample. Our objective will be to obtain empirical validation of the system, and test it.

We will decide how to implement it, depending on the broker/exchange, speed requirements, etc., and tests will be initialized in a real but controlled environment over the procedures, for future live implementation.

At this point, the researcher must evaluate the validity of the strategy, and judgment and experience are vital.

Tips for Developing Models

• Select your initial research sources well. Avoid gurus, or any other initial source that lacks scientific reasoning.
• Understand the ideas of what you want to investigate, the premises behind the model, what decisions it makes, and why it makes them
• Formulate alternative hypotheses. Your main objective is to discover that your model is so robust it cannot be "broken." But it is always advisable to have alternative hypotheses.
• Make (reasonable) changes to the initial premises, if the premises or findings suggest it, in order to increase robustness
• Look for evidence to the contrary, to strengthen the main hypothesis
• Focus on causality and probabilities rather than statistical correlation.
• Decompose the model into parts or objects, and try to validate them individually.
• Know yourself. Find out how much you know about the investment thesis, and what things you know you don't know. (things you DON'T know you DON'T know fall within the random behavior of the market)

$$\text{IR} = \frac{\mathbb{E}[R_p - R_b]}{\sigma_{tracking}}, \quad \text{IC} = \rho(\hat{R}, R)$$

$$H_0: \mu = 0, \quad t = \frac{\bar{R}}{s / \sqrt{N}}, \quad p = 2 \times P(T > |t|)$$

$$S = \frac{\mathbb{E}[R_p - R_f]}{\sigma_p}$$

Use of Models

The uses of models change depending on who is using them. For example, for quantitative investors, models produce signals about expected future returns on an asset or portfolio, while for fundamental investors, a model's forecast creates new information that they need to synthesize along with other qualitative information, in order to improve decisions.

The model itself has no ultimate general purpose, other than providing information and advantages to those who use it when obtaining future returns. The key to success lies in synthesizing information from multiple sources, finding the strengths and weaknesses of each approach.

Unlike technical analysis, which we could define as a group of people trying to predict the direction of a vehicle driven by a drunk monkey based on the marks the tires leave on the asphalt. And through an accumulation of biases, self-sabotage their logic and validate their trading.

References

[2016] Best practices in research for quantitative equity strategies. The Journal of Portfolio Management 42 pp 135-143