Glossary

Plain-English definitions of the terms used on this site, with a link to where each one appears.

A B C D F G H I K L M N O P R S T V W

Arena: One pool of stocks plus the index fund we compare it with: S&P 500 against SPY, Nasdaq-100 against QQQ, small caps against IWM. Where you'll see it: Evolution runs
Base rate: How often something happened across comparable past cases. If a strategy beat SPY in 4 of 10 past months, its base rate for "beats SPY" is 40%. It is where every forecast starts. Where you'll see it: Forecasting primer
Benchmark: The plain index fund a strategy has to beat to be worth anything: SPY, QQQ or IWM, depending on the arena. Where you'll see it: Evolution runs
Breeding: How new strategies are made from good ones. Two settings cards can be mixed (crossover), and some values are nudged at random (mutation). Evolution never writes code. Where you'll see it: How it works, evolution
Brier score: The squared miss of a probability forecast: (probability minus outcome) squared. 0 is perfect, always saying 50% scores 0.25, and lower is better. Where you'll see it: Forecasting, Brier score
Calibration: Whether the things we call 30% likely happen about 30% of the time. A calibration chart plots stated odds against how often each came true. Where you'll see it: Forecasting, calibration chart
Champion: The strategy a run ranks first after its neighbor check. It is the one traded on the unseen days and reported. Where you'll see it: Evolution runs
Coin-flip twin: For every forecast we issue, a twin question gets a flat 50%. Beating the twin's Brier score is the minimum bar for having skill. Where you'll see it: Forecasting, why a coin flip
Concentration (top-2 share): How much of a strategy's gain came from its two biggest winners. Above 50% means the result rests on one or two stocks, which is one bet, not a pattern. Where you'll see it: Reading the reports
Deflated Sharpe ratio: A Sharpe ratio (return per unit of risk) marked down for how many strategies were tried before this one was picked. Above 0.95 would be convincing; the more tries, the harder the bar. Where you'll see it: Reading the reports
Drawdown: The fall from a high point to the next low point, as a percentage. A max drawdown of 25% means paper money once sat 25% below its best level. Where you'll see it: Reading the reports, scoreboard columns
Fair random portfolios: The luck test. At least 100 portfolios of 5 random eligible stocks, fully invested and rebalanced on the champion's schedule, trade the same unseen days. The champion's percentile among them says how much luck alone could do. Where you'll see it: How it works, luck checks
Fitness: The training score used to pick survivors: the average fold score minus half its spread, so steady beats lucky. Each fold is measured against textbook momentum. Where you'll see it: How it works, evolution
Fold: One slice of the training period. Each strategy is run on every fold separately, so it has to do well in several stretches and not only one. Where you'll see it: How it works, evolution
Forward test: Strategies frozen on a start date that then trade paper money every weekday on prices nobody had yet. The only evidence that cannot be fitted to the past. Where you'll see it: Forward test
Genome: The settings card behind a strategy: 21 values such as how many stocks to hold and how often to trade. A bot just follows its card. Where you'll see it: The genome
Hindsight members: A stock list built from today's index members. It leaves out companies that were dropped or went bust, so old results look too good. We tag runs that used one. Where you'll see it: Evolution runs
Holdout: The last stretch of days, about six months, kept out of training. The champion trades it once, at the end. A result there is honest in a way a training result cannot be. Where you'll see it: How it works, holdout
Information coefficient (IC): How well a ranking signal predicted next month's stock returns, as a rank correlation. Near 0 means no predictive power; a t-statistic beyond 2 is the usual bar. Where you'll see it: What we've learned, signal IC
Island: A separate population that evolves on its own and trades its best bot with the others every few generations. It keeps the search from collapsing onto one idea. Where you'll see it: How it works, evolution
Kill switch: A strategy stops trading if its paper money falls 35% from its peak. The forward test records the stop. Where you'll see it: Forward test rules
Look-ahead bias: Using information the strategy could not have had at the time, such as tomorrow's price. We defend against it by deciding after the close and filling at the next open. Where you'll see it: How it works, data
Luck check: The set of comparisons that ships with every headline number: percentile among fair random portfolios, return against textbook momentum, deflated Sharpe, concentration, and whether the period was unseen. Where you'll see it: How it works, luck checks
Multiple testing: Try enough strategies and one will look great by chance. The deflated Sharpe ratio is the correction for this. Where you'll see it: How it works, luck checks
Neighbor check: A finalist is re-run with six slightly changed copies, each moving one setting. Its rank is the lower of its own score and the median of its neighbors, so a lucky spike on one exact setting does not win. Where you'll see it: How it works, evolution
Overfitting: Fitting a strategy to the quirks of past data so well that it fails on new data. Evolution is very good at it, which is why the unseen days exist. Where you'll see it: How it works, holdout
Paper money: Simulated money in our own ledger. Nothing here places a real order; Alpaca supplies prices only. Where you'll see it: Start here
Percentile: Where a result ranks among fair random portfolios on the same days. 58 means it beat 58% of them; 50 is average luck, and under 95 is consistent with luck. Where you'll see it: Reading the reports
Point-in-time membership: Using the index members as they were on each past date, joiners and leavers included, instead of today's list. It removes survivorship bias. Where you'll see it: How it works, data
Rebalance: Selling and buying so holdings match the strategy's plan again. Most genomes rebalance about monthly (every 21 trading days). Where you'll see it: The genome
Shake-up: When an island has not improved for 3 generations, its next children are mutated hard and a third are replaced by random newcomers. Where you'll see it: How it works, evolution
Sharpe ratio: Average daily return divided by how much it bounces around, scaled to a year. Higher means a smoother ride per unit of gain. Where you'll see it: Reading the reports
Slippage: The small cost charged on every trade because real fills are worse than the quoted price: 0.02% for funds, 0.05% for big-index stocks, 0.20% for small caps. Where you'll see it: The genome
Superforecasting: The method from the 2015 book by Philip Tetlock and Dan Gardner: precise probabilities, start from base rates, keep score, learn from misses. Where you'll see it: Forecasting primer
Survivorship bias: Testing only on companies that still exist. The failures are missing, so the average looks better than it was. Where you'll see it: How it works, data
Textbook momentum (12-1): Buy the 10 stocks that rose most over the past year, skipping the latest month. It is our standard comparison: an evolved strategy that cannot beat it has found nothing new. Where you'll see it: Strategy gallery
Verdict: The one-line label on a scoreboard row, such as "no signal yet: champion within the range of fair random portfolios". It is written by rule from the luck check, not by hand. Where you'll see it: Reading the reports
Walk-forward: Repeat the whole test on several half-year windows: evolve only on data before a window, then trade the window blind. Results across windows say more than one run. Where you'll see it: How it works, walk-forward