Algorithmic trading : winning strategies and their rationale

If you've ever coded a strategy that looked brilliant in a backtest and bled money live, this book explains why. Chan, a quant practitioner, shows the full pipeline — statistical tests, execution, risk — with real numbers from real ETFs and futures. It's for traders ready to move from hunches to systems, and for anyone who wants to see how professionals separate signal from curve-fit noise.

1. 1Chapter 1

   A backtest is your only honest referee — if you don't cheat it

   Picture a trader staring at a beautiful equity curve, ready to wire real money. Chan insists a backtest is not a guarantee of future performance, but it is the only objective way to judge a strategy before committing capital. Survivorship bias in databases like Compustat inflates small-cap backtest returns by 1.5% a year and large-cap by 0.5%. Look-ahead bias — using today's close to decide today's entry — can make a losing strategy look profitable. And transaction costs around 0.5% per trade for a typical retail algorithmic trader can flip a seemingly profitable high-frequency strategy into a loser.

   The idea

   The backtest's value depends entirely on its integrity. Three quiet errors — missing dead companies, peeking at future prices, ignoring costs — each push results in the same flattering direction, so a flawed test doesn't just add noise, it adds systematic optimism. That's why a strategy that survives a clean backtest is rare and valuable, and why the trader's first skill isn't finding signals but auditing their own testing pipeline for self-deception.

   Try this: Before trusting any backtest, check three boxes: does the data include delisted stocks, could the entry rule have known its own input at decision time, and are per-trade costs subtracted?

2. 2Chapter 2

   Mean reversion starts with a statistical lie detector

   Chan brings in the Augmented Dickey-Fuller test, the standard statistical check for mean reversion: if the test statistic is more negative than the critical value of -3.5, you have mean reversion at the 95% confidence level. Then he puts it to work on EWA, the iShares MSCI Australia ETF, with a Bollinger Band strategy: enter on a move 2 standard deviations below the rolling 20-day mean, exit at the mean. From 1996 to 2007 that simple rule earned a Sharpe ratio of 1.1. He also introduces the half-life of mean reversion, -log(2)/log(lambda) from a regression coefficient, and says a half-life of 5 to 20 days is practical for trading.

   The idea

   Mean reversion is a measurable property with a clock. The ADF test tells you whether the wobble is real; the half-life tells you how fast prices snap back, which sets your holding period. A half-life outside the 5-to-20-day window means the reversion is either too slow to harvest efficiently or too fast for daily data to catch. Matching your trade duration to the measured half-life is what turns a statistical curiosity into a strategy.

   Try this: Run an ADF test on a price series you trade, then compute its half-life — if it lands outside 5 to 20 days, reconsider your holding period.

3. 3Chapter 3

   The gold pair that printed money — then broke

   Chan describes a pairs trading strategy on GLD, the SPDR Gold Trust, against GDX, the VanEck Gold Miners ETF, from May 2006 to December 2007. From May 2006 to December 2007 the strategy annualizes 33.3% with a Sharpe ratio of 2.0. The GLD-GDX pair stopped being cointegrated after 2007, and a trader who kept going suffered a drawdown of over 20% in 2008. Chan's broader answer is the Johansen test, which finds cointegrating relationships among multiple assets — he applies it to a basket of four energy ETFs, XLE, OIH, USO, and UNG, from 2007 to 2009, achieving a Sharpe ratio of 1.5.

   The idea

   Cointegration is a statistical relationship, not a law of nature — and it has an expiry date. The same economic logic that made GLD and GDX move together can quietly dissolve when the underlying businesses diverge. The practical defenses are diversification across multiple cointegrated relationships, found by tests like Johansen's, and a standing assumption that any pair's stability is temporary.

   Try this: If you trade a pairs strategy, schedule a recurring cointegration re-test — and predefine the exit if the relationship fails it.

4. 4Chapter 4

   Cross-sectional mean reversion: buy yesterday's losers, minus the costs

   Chan reports a cross-sectional mean reversion strategy on S&P 500 stocks — buy the 50 stocks with the lowest previous day's returns, short the 50 with the highest — yielded an average annual return of 7.2% from 1996 to 2007 before transaction costs. With a 0.1% one-way transaction cost, the return drops to 2.3% annually.

   The idea

   High-turnover strategies live or die on a number most backtests treat as an afterthought. When you trade the entire index daily, even a seemingly tiny 0.1% one-way cost compounds into the dominant force on returns. The gap between 7.2% and 2.3% is the true price of frequency — and it explains why strategies that look mediocre on paper but trade rarely can beat flashy ones that trade constantly.

   Try this: Take your strategy's annual turnover, multiply by your realistic one-way cost, and subtract that from the backtest return before judging it.

5. 5Chapter 5

   Momentum pays — until the crash

   Chan presents a time-series momentum strategy on S&P 500 futures: go long when the 12-month return is positive, short when negative. From 1980 to 2009 that strategy annualized 10.5% with a Sharpe ratio of 0.6. Jegadeesh and Titman's 1993 study, where buying the top decile of stocks by past 6-month returns, shorting the bottom decile, and holding 6 months yielded an average monthly return of 1% from 1965 to 1989. Momentum strategies can suffer sudden crashes: in 2009, a long-short momentum portfolio lost 73% in three months, from March to May.

   The idea

   Momentum is one of the most durable effects in finance, documented across decades and markets — but its returns come with embedded crash risk. A Sharpe ratio of 0.6 already suggests modest steady gains punctuated by rare, severe losses.

   Try this: Before adding any momentum strategy, compute its worst historical three-month loss and size positions so that loss wouldn't force you to quit.

6. 6Chapter 6

   The overnight drift and the opening gap

   Chan finds two simple patterns in SPY, the SPDR S&P 500 ETF. First, momentum: if the previous day's close was above the open, buy at today's open and sell at the close. From 2000 to 2009, that yielded an average daily return of 0.04%. Second, mean reversion: when the opening gap exceeds 1 standard deviation of the 20-day average gap, fade it. That gap-fading strategy produced a Sharpe ratio of 1.2 over the same decade.

   The idea

   Intraday markets have structure that changes with the size of the move. The same ETF obeys both rules because they trigger under different conditions.

   Try this: Log SPY's opening gap against the 20-day average gap for two weeks and note which days exceed 1 standard deviation — watch how often the gap closes by the bell.

7. 7Chapter 7

   Factor models: separating skill from the market's tide

   Chan brings in the Fama-French three-factor model, where a stock's expected return is E(R) = Rf + beta*(Rm-Rf) + s*SMB + h*HML, and uses it to hedge the market exposure of a long-short portfolio. After hedging, the portfolio's annualized alpha from 1996 to 2007 was 4.8%.

   The idea

   Factor regression strips out the exposures anyone could get cheaply — market beta, size, value — and shows what remains. Hedging the known factors concentrates your risk where you actually have an edge.

   Try this: Regress your strategy's monthly returns on market, size, and value factors to see how much of your performance is just beta wearing a costume.

8. 8Chapter 8

   The Kalman filter: a hedge ratio that adapts

   Chan applies the Kalman filter to dynamically estimate the hedge ratio for the EWA-EWC pair — iShares MSCI Australia against iShares MSCI Canada. From 1996 to 2007, the adaptive strategy annualized 12.4% with a Sharpe ratio of 1.4.

   The idea

   The Kalman filter treats the hedge ratio as a hidden state that evolves, updating with each observation. That adaptability improves performance across a decade-long span.

   Try this: If you run a pairs trade, replace the fixed hedge ratio with a rolling regression or Kalman estimate and compare drawdowns over the same period.

9. 9Chapter 9

   Machine learning finds patterns — including the fake ones

   Chan feeds a support vector machine 10 technical indicators and asks it to predict the next day's direction of the S&P 500. From 2000 to 2009, the SVM achieves 55% accuracy. The strategy, though, was not profitable after transaction costs. Neural networks and SVMs are prone to overfitting, and a model with 100 parameters trained on 1,000 data points will almost always find a pattern in noise.

   The idea

   With enough flexible parameters, fitting the past is guaranteed — the model will memorize noise and present it as insight. A 55% directional edge sounds real until costs eat the thin margin it supports.

   Try this: Count your model's parameters and your training observations — if the ratio approaches 1:10, cut features or gather more data before trading it.

10. 10Chapter 10

    Selling volatility: steady income with a hidden cliff

    Chan examines the volatility risk premium: sell a 1-month at-the-money straddle on the S&P 500 and delta-hedge daily. From 2000 to 2009, this yielded an annualized return of 8.3% with a Sharpe ratio of 1.1. The strategy lost 30% in October 2008.

    The idea

    The volatility risk premium is real — implied volatility tends to exceed realized — but harvesting it means selling insurance against market panics. A decade of 8.3% annual returns can coexist with a 30% single-month hit.

    Try this: For any income-style strategy you run, compute its worst single month historically and ask whether your position size survives it without forced exit.

Traders ready to replace intuition with tested systems, and quants who want a practitioner's view of the full strategy pipeline. Comfort with basic statistics and some programming will help you get the most from it.

FAQ