Key Takeaways
Treat the backtest and live trader as one system, not two projects.
Make research executable. An algorithm takes historical or live market data, generates signals, and sends orders through a broker. Chan’s practical rule is to keep the backtest and live system as close as possible, ideally using the same trading logic. A strategy that works in a research script but behaves differently in production has not really passed its test. Data quality matters just as much: survivorship bias can omit failed or delisted companies, while ordinary consolidated opening and closing prices can differ from the auction prices a trader could actually receive. Options data should include closing bids and offers, not just the last trade, which may be stale.
Audit the whole chain. Historical data, broker reliability, fees, order execution, and software bugs all affect results. Backtests are prototypes, not ready-made products.
The book’s systems view resembles reliability engineering: a chain is only as dependable as its weakest component. A signal can be statistically excellent yet commercially useless if its price feed is stale or its orders fail to fill. This also clarifies why reproducibility matters. Using identical logic in research and production reduces one source of discrepancy, but it does not make historical data accurate or execution predictable. Treating data provenance, software versioning, and order records as part of the model is a stronger standard than merely checking whether a simulated equity curve rises.
Judge a strategy by its drawdown path, not Sharpe ratio alone.
Survival is a metric. CAGR measures compounded growth, while Sharpe ratio compares average return with return volatility. Both can mislead when returns have unusually large losses. Chan favors the Calmar ratio, which divides CAGR by the absolute size of the worst drawdown over the most recent three years. A strategy can look attractive on average and still be impossible to endure when it loses sharply.
Leverage magnifies assumptions. The Kelly formula estimates a growth-maximizing allocation from expected return and variance, but it assumes a simplified return distribution. In Chan’s example, a strategy with mean daily excess return near 0.05 percent and volatility near 1 percent suggests leverage of five. A single loss like the roughly 20 percent SPY decline on Black Monday could wipe out that leveraged account. He recommends scaling down until crisis-period drawdowns are tolerable.
This is a useful distinction between statistical risk and lived risk. Volatility treats gains and losses symmetrically, but investors experience a deep loss differently from an equally large gain. Maximum drawdown better captures that asymmetry, though it is still only one historical observation and can understate future extremes. The practical insight is not that one ratio is universally superior, but that no single summary statistic can stand in for a loss distribution, liquidity needs, and an investor’s capacity to stay invested.
Expect factor returns to survive because their risks cannot be diversified away.
Know what pays you. A factor model relates returns to common drivers. A factor risk is exposure to a systematic force, such as the broad market, value, momentum, or volatility, that remains even when a portfolio holds many securities. By contrast, idiosyncratic noise can shrink through diversification. Chan argues that factor returns can persist because investors demand compensation for bearing these harder-to-escape risks, while pure alpha is scarce and can be crowded or capacity-limited.
Prediction is not explanation. A conventional model may explain why a stock moved without forecasting its next move. In one test, Fama-French market, value, and size factors failed to predict next-day S&P 500 stock returns out of sample. Other factors performed differently across universes: fundamental signals that looked promising in small-cap stocks did not translate cleanly to large-cap, market-neutral trading.
The distinction between a risk premium and an anomaly is central to interpreting a backtest. A systematic return can persist precisely because it is uncomfortable to hold through bad periods, not because the market has overlooked a simple trick. That makes exposure management as important as prediction. The caveat is that a historical premium does not guarantee future compensation: preferences, regulation, and market structure can change. Factor labels describe hypotheses about shared risk, not permanent laws of nature.
Prefer simple models until clean out-of-sample tests justify added complexity.
Models need restraint. Chan demonstrates autoregressive models, which predict a series from its own past, vector autoregressions for related assets, and Kalman filters that estimate changing hidden quantities such as a hedge ratio. He also surveys machine-learning tools, including regression trees, support vector machines, and neural networks. These methods can find patterns, but financial returns shift over time and offer fewer reliable examples than many machine-learning tasks.
Test what generalizes. On SPY, ordinary regression produced a 0.4 percent out-of-sample CAGR, while stepwise regression, which selects a smaller set of predictors, produced 10.6 percent in the same historical test. More elaborate neural networks often fit training data better but failed to improve future-period results. When combining stocks, predictors and returns must be normalized so that one volatile stock does not dominate the model.
The results illustrate a broad principle of statistical learning: flexibility reduces bias but increases variance, especially when the sample is small and the data-generating process changes. A further caution is that randomly dividing time-series observations for cross-validation can put near-neighbor periods in both training and validation sets, making performance look more stable than it is. Chronological tests and repeated regime checks are often more revealing. Even then, a modest Sharpe ratio or a positive CAGR is evidence to investigate, not proof of a durable edge.
Sell volatility only when the premium compensates for catastrophe risk.
Options are not pure volatility bets. A short option position collects premium and benefits from time passing, but it can lose heavily when prices move sharply. In Chan’s historical comparison, shorting and rolling the front volatility future produced a 17.8 percent CAGR, compared with 7.2 percent for holding SPY, but the short-volatility strategy also suffered a maximum drawdown of about 92 percent. The return came with a severe obligation.
Separate realized from implied. A GARCH model forecasted the direction of next-day realized volatility correctly more than two-thirds of the time in one SPY test. Yet that did not mean volatility-linked products would rise when realized volatility rose. Their prices also reflect option premiums and time decay. Other approaches, such as gamma scalping, pair a mean-reversion trade in the underlying with long options as a tail hedge; dispersion trades exchange index-option exposure for individual-stock options and correlation risk.
The insurance analogy is powerful because it makes the asymmetry visible: small, frequent premium gains can coexist with rare, enormous claims. A high backtested return is therefore inseparable from the financing and survival plan for crisis periods. The distinction between realized and implied volatility is equally important. A forecast can be accurate about one quantity and still fail as a trading signal for an instrument priced on several quantities. Strategy design should start with the risk being sold, not the headline yield.
At intraday horizons, execution quality can become the entire strategy.
The quote is not the fill. Intraday trading makes small profits, so bid-ask spreads, slippage, market impact, latency, and missed orders can erase an apparent edge. An order book shows available bids and offers, but a limit order’s fill also depends on its queue position. Dark pools may offer execution near the midpoint, yet their prices and resting orders can expose traders to latency advantages or adverse selection.
Order flow carries information. Order flow is signed trading volume: buyer-initiated trades add positive volume, and seller-initiated trades add negative volume. It can help forecast short-term price changes, but a backtest must model how orders actually execute. In one ES futures illustration, an order-flow strategy made about $300 using market-order assumptions, versus $762.50 assuming midpoint fills. The more flattering result ignored the reality that urgent trades usually cross the spread.
This chapter treats execution as part of the strategy rather than a final administrative step. That is especially important in fragmented markets, where the visible best quote may be small, stale, or unavailable by the time an order arrives. The order-flow example also shows how a realistic fill model can reverse the apparent attractiveness of an idea. Still, execution is not merely friction: order placement can change queue priority, reveal information, and affect market impact. The simulation must represent those interactions, not just subtract a fixed fee.
Treat bitcoin arbitrage as credit-risk trade, not a free lunch.
New markets reward familiar tests. Chan applies time-series models, mean reversion, machine learning, and order-flow signals to bitcoin, a market that traded continuously and had large price swings. A Bollinger-band strategy, which buys when price falls sufficiently below a moving average and sells after it returns, showed a 42 percent CAGR in one historical test. An autoregressive model produced far more spectacular results, but depended on unrealistic assumptions about instant midpoint fills. These figures describe a past simulation, not a dependable forecast.
The spread has a reason. A quoted bitcoin price gap between exchanges may appear to offer arbitrage. In Chan’s example, the gross gap was $5.64 per coin. Commissions and transfer costs reduced it, while the cheaper exchange’s weaker creditworthiness created a risk that funds could be lost. The book cites a study estimating that 45 percent of bitcoin exchanges failed through thefts or hacks.
Bitcoin serves as a clean demonstration of a recurring error: mistaking a price discrepancy for a riskless opportunity. The order may be profitable only if both venues honor withdrawals, settlement occurs on time, and execution matches the observed quotes. Those operational conditions are part of the asset’s economics. The comparison also highlights the value of transferring methods across markets while refusing to transfer assumptions. A model that worked on a liquid exchange-traded asset does not carry its execution quality or institutional protections into a young exchange ecosystem.
Keep trading strategies replaceable, since profitable edges have finite shelf lives.
Strategies age. A once-profitable system can weaken when market structure or economic conditions change, or when other traders exploit the same opportunity. Chan illustrates this with a GLD-GDX pairs strategy: the spread shifted for years, hurting the backtest, then stabilized, making conditions more favorable again. A different GLD-USO strategy continued to perform in the period described. One failed test need not mean every related idea is dead, but it demands diagnosis rather than blind loyalty.
Diversify by independent ideas. Chan recommends maintaining at least ten live strategies, with continuing creation and retirement. As a strategy loses money, its allocation can be reduced; stronger or newer strategies can take its place. This portfolio-level view limits dependence on any single signal and acknowledges that strategy research is an ongoing process, not a search for one permanent machine.
A portfolio of strategies resembles a portfolio of scientific hypotheses: each has a domain of validity, and evidence should update confidence rather than trigger either permanent belief or instant rejection. Diversification helps only when the strategies are genuinely distinct. Ten variations of the same momentum exposure can fail together during a regime shift. The useful discipline is to track not just returns, but the mechanism each strategy relies on, its shared risks, and whether observed losses are ordinary noise or evidence that the mechanism has changed.
Pursue trading for the research and service, not easy wealth.
Profit is not the whole case. Chan presents algorithmic trading as intellectual work at the intersection of mathematics, computing, and economics, while warning that most traders do not profit. He also argues that independent research can offer autonomy and flexibility, including the practical benefits he associates with working from home. These are personal advantages, not guarantees of health, income, or freedom from stress.
Trading can provide a service. Market makers and traders can supply liquidity to investors who need to transact, while an investment manager can help safeguard retirement or institutional assets. The book also describes teaching, consulting, writing, and trading as mutually reinforcing activities: explaining methods can sharpen research, and relationships built through useful work can create collaboration. Managing other people’s money adds responsibilities around transparency, structure, risk, and conflicts of interest.
This closing argument broadens the definition of a worthwhile career beyond salary, but it remains grounded in a high-risk occupation. Autonomy can be valuable, yet self-employment also shifts responsibility for income, compliance, and emotional resilience onto the trader. The service claim is strongest when linked to concrete functions such as liquidity provision and accountable investment management, rather than treating all trading as inherently socially useful. The book’s most defensible point is that intellectual curiosity and useful relationships can make a volatile business more sustainable.
Analysis
Machine Trading is best read as a field guide to turning financial hypotheses into systems that can survive contact with actual markets. Its chapters move from data and infrastructure to factors, time-series models, machine learning, options, intraday mechanics, bitcoin, and the trader’s working life. That breadth is also its summarizing challenge: the value lies less in one master formula than in the repeated habit of testing assumptions at every layer.
The book’s most consequential claim is methodological. A backtest is not a verdict. It is a conditional result produced by particular data, instruments, execution assumptions, and historical regimes. The examples make this concrete: apparently powerful strategies weaken when they are made market-neutral, tested on large-cap stocks, or charged realistic spreads. Extreme returns in bitcoin and volatility products are presented alongside the caveats that make them difficult to realize. This tension is the book’s strongest teaching device.
The framework is practical, but the evidence needs careful reading. Many reported results come from specific historical windows, and model selection itself can create false confidence when many variants are tried. Out-of-sample testing reduces this danger but cannot eliminate regime dependence or guarantee tradability. Likewise, diversification among strategies works only when their underlying risks are meaningfully different. These limitations do not undermine the book’s toolkit; they define how responsibly to use it.
The enduring lesson is to treat trading as an iterative engineering discipline. Data integrity, model simplicity, risk sizing, execution, and strategy retirement are not side issues. They are components of the same decision system. The reader should come away less impressed by spectacular simulated returns and more interested in what assumptions generated them, what could break them, and whether the result remains credible after costs and operational risks are included.
Glossary
Calmar ratio
Growth relative to drawdownA risk-adjusted performance measure calculated by dividing compound annual growth rate by the absolute value of the maximum drawdown over a specified period. Chan favors measuring drawdown over the most recent three years, so the statistic is less dependent on the full length of a backtest than the related MAR ratio.
Kelly formula
Growth-maximizing leverage estimateA formula for choosing the allocation or leverage that maximizes expected long-run compound growth under specified assumptions about returns. In the book, the calculation is based on estimated return and variance. Chan warns that its assumptions can make the recommended leverage too aggressive, particularly when returns have severe losses not captured by a Gaussian model.
Factor risk
Exposure to shared return driversRisk arising from exposure to a common source of return, such as the market, value, momentum, or volatility. Unlike stock-specific noise, factor risk does not disappear simply by adding more securities with similar exposure. Factor models help explain, forecast, or manage these shared exposures, though a factor’s historical return is not guaranteed to persist.
Cross-validation
Model testing within training dataA procedure that divides a training sample into subsets, fits models on some subsets, and evaluates them on the omitted data. Repeating this process allows comparison of models using data not used in each individual fit. It can help detect overfitting, although random folds may be misleading for time series when adjacent observations share market conditions.
GARCH
Model of changing return volatilityGeneralized autoregressive conditional heteroskedasticity, a family of models in which predicted return variance depends on previous predicted variances and past squared returns. Chan uses it to forecast the size of future returns. A forecast of realized volatility does not automatically predict the price of volatility-linked products, which also reflect implied volatility, time decay, and other variables.
Gamma scalping
Hedge options with underlying tradesAn options strategy that holds a long straddle or strangle and trades the underlying as its price moves away from a reference level. The options’ gamma changes the portfolio’s delta, creating opportunities to rebalance through mean-reversion trades. The options can protect against extreme moves, but their time decay can outweigh trading gains when the underlying moves too little.
Bulk volume classification (BVC)
Estimate signed volume from barsA method for estimating buy and sell order flow from price changes and volume in aggregated bars, rather than classifying each individual trade. It uses the distribution of price changes to estimate the fraction of volume initiated by buyers or sellers. The method is less data-intensive than trade-by-trade tagging, but the book notes that its estimates can be less accurate.
Order flow
Signed volume of initiated tradesA measure of trading volume that assigns positive signs to buyer-initiated transactions and negative signs to seller-initiated transactions. Aggregated over a time interval, it can help indicate short-term buying or selling pressure. Accurate calculation requires trade-direction information, such as an exchange aggressor tag, or an estimate from prices and volume.
Adverse selection
Loss from informed counterpartiesThe risk that a passive order is filled just before the market moves against it because the trader on the other side has better short-term information. A resting buy order, for example, may be executed when informed sellers anticipate a price decline. This risk affects liquidity-providing orders and can make a strategy’s simulated fills look more profitable than its realized trades.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.