Introduction to quantitative trading

Mislav Šagovac

Terminology

• Automated (algo) vs manual - automating the process, no human in a loop.
• Systematic vs discretionary - using a system (rules) for high level decisions.
• Quantitative va non-quantitative - using data and statistical models to make trading decisions.

Mostly focus on equity stocks, but most of the concepts can be applied to other asset classes. Mostly applied for retail investors. There is of course semiautomated trading. A strategy which watches prices on two different exchanges and trades whenever there is a profitable arbitrage isn’t really quantitative, although it is systematic and automated. Similarly some approaches to rules-based trend following (eg buy if the 20 day MA is above the 200 day) are not really quantitative, even though they are systematic and automated.

Generate ideas

• SSRN - repository of academic research papers in finance and economics. Look at G1, C58 JEL codes.
• QuantConnect - QC has list of trading ideas
• Robot Wealth blog and bootcamp - is a blog and online course platform that provides resources on quantitative trading strategies.
• Quant roadmap from X user quant_arb. This roadmap provides detailed information on how to get started with quantitative trading
• Quantpedia
• Social media - Twitter and reddit.

Tips for ideas

List of strategies you can start with.

Example of strategies

• Momentum
• Trend following
• Short volatility
• Statistical arbitrague (Pairs trading)
• Factor analysis
• Event-based - PEAD, PREAD, M&A, index rebalancing
• ML driven strategies

Data

Security Master

• corporate actions:
  • stocks splits
  • dividends
• symbol changes
• delistings
• mergers and acquisitions
• Vendors: Quantconnect, Databento, Bloomberg

Market Data Types

• MBO (Market by Order): Full order book, L3.
• MBP-10 (Market by Price): Market depth, L2.
• MBP-1 (Top of Book): Trades and quotes, L1.
• TBBO (Top of Book, Sampled): Sampled in trade space.
• Trades: Tick-by-tick trades, last sale.
• OHLCV-T: Aggregates per second, minute, hour, or day.
• time bars, tick bars, volume bars, dollar volume bars.

First 3 are mostly used for HFT, but it can be used for higher frequency, for example for generating predictors.

Pricing data providers

• Historical data vendor list
• Equities
  • If you have money: Bloomberg, Thomson Reuters, Morningstar
  • If you have some money: Databento, QC bulk OHLCV, FMP cloud, Norgate Data, IB API, Algoseek
  • If you have no money: Yahoo Finance, Ducascopy, investing.com (scraping).
• Crypto
  • Exchanges: Binance, Coinbase
  • Kaggle datasets.
• Government US yields:
  • Liu and Wu (2020)
• Options data
  • Databento, ORATS

Fundamental data

• Equities
  • Financial reports, SEC filings, 8-K, analyst estimates
  • analyst ratings, insider trading, short interest
  • If you have some money: FMP prep, SimFin+
• Economic data
  • FRED, BIS, World Bank, OECD

Alternative data

• News and social media
  • News API - Tingo
  • Twitter - look at Rapid API.
  • Reddit - look at Pushshift API.
• Weather data - for commodity trading
• Other - satellite images, shipping data, credit card transactions.

Data Quality

• Survivorship bias - only data from companies or ETF’s that still exist.
• Look-ahead bias - using data that would not have been available at the time of the trade.
• Restated data - data that has been changed after the fact (earnings, GDP).
• Data cleaning
  • duplicates
  • missing values - lcfo,
  • outliers - winsorization .

Data tips

• Spend some money - for high quality data.
• Data storage - flat files -> parquet database.
• Visualize your data - before you start modeling.

Predictors

Predictors

• Technical indicators - moving averages, RSI, MACD.
• Statistical indicators - volatility, skewness, kurtosis.
• Sentiment indicators - Twitter, Reddit, news.
• Economic indicators - GDP, unemployment, inflation.
• Fundamental indicators - P/E, P/B, EPS.
• Alternative data - satellite images, shipping data, credit card transactions.

What is a good predictor?

• Predictive power - correlation with future returns.
• Portfolio sorts - sort stocks based on predictor and see if it (lineary) predicts future returns.
• ML models - use ML models to predict future returns and check performance out-of-sample.
• Feature importance - check which features are most important (some models have this feature).
• Brute force - try all predictors and see which one works best on backtest (dangerous).

Predictors tips

• Stationarity - stationary predictors have some good properties.
• Ensemble - use ensemble of various predictors.
• Feature engineering - create new predictors from existing ones (various transformations).

Tools & Platforms

Programming languages

• Python - most popular. Packages: Pandas, NumPy, SciPy, Scikit-learn, TensorFlow.
• R - great for research and statistics, quantmod, TTR, xts, data.table, tidyverse.
• other: Julia, Matlab, Go.
• for speed: C++, Rust.

Research tips

• visualize - before you start modeling (scatter plots, bars…).
• parameter sensitivity - check how the model behaves with different parameters values.
• sample splits - walk forward, cross-validation, Monte Carlo simulations.
• robustness - robust to different market conditions, years, asset classes etc.
• causal effects - what explains the effect?

Backtesting engines

• QuantConnect (py, C#) - my recommendation; free data in web development! All assets classes.
• Backtrader (py) - open-source, no heavy dependencies, good docs.
• Zorro (c, c++) - vary fast, py and R integration.
• Nutilius Trader (py, rust) - relatively new, but promising, fast, crypto.
• ctxx (py) - for crypto, haven’t tried, but heard good things.
• VectorBT - for vectorized backtesting.
• Build yours - not in the beggining.

Backtesting pitfalls

• Overfitting - tuning the strategy to the data, p-hacking.
• Look ahead bias - we mention here again.
• Only insample - training data, out of sample - testing data.
  • Walk forward optimization
  • Cross-validation
  • Montecarlo simulation

Backtesting tips

• Backtesting is not a research tool; it should be used at the end of the process:
  1. Final Verification - don’t optimize parameters.
  2. Sensitivity - to transaction costs and fill ratios.
  3. Code bugs
  4. Non-Performance Stats: turnover, max net/gross exposure, and other stats.
• Avoid Full Backtests for Signal Evaluation: Use alternative methods like event studies or regression for signal/alpha evaluation without running full backtests.

Costs - Reality modeling

• Brokerage costs - easily estimated, commissions and fees.
• Trading costs - slippage (delay), bid-ask spread and market impact.
• Opportunity costs - order fillings.
• Short sale constraints - cost of borrowing.
• CFD costs - if you trade CFD’s

Example

Example: Momentum strategy for stocks equity

• Difference between momentum (relative or cross sectional momentum) and trend following (absolute ot time series momentum):
• Trend - moving in the same direction. If it has gone up rcently it will continue to go up.
• Momentum - if it has gone up relative to other assets, it will continue to go up.
• we will follow this analysis.

Causes of momementum / trend

• Behavioral - Overreact/underreact to new information, anchor to old prices, herding.
• Structural - Investors are slow - need to call investment committee etc.
• Information-based: Information diffusion, information asymmetry.
• Trend effects are are stronger in small caps, in countries with less developed financial markets, after important information is released or fair value is less clear in general.
• Crypto - fragmented, hard to value, strong retail engagement.

Some code

• Research with R and quarto
• Backtest in Quantconnect

Momentum tips

• long term momentum behaves more like factors (above week).
• there are explosive momentum after shocks, but they are hard to trade due to low capacity.
• use as many markets as possible.
• ensemble
• lower the frequency, longer the backtest.

Thank you!