Reading a live strategy

A house strategy page shows two records: the backtest before launch, and the live record since launch. They are not the same kind of evidence, and only the second is a claim.

Backtest before launch, record since launch

The rules of every strategy were chosen by people who knew how the previous decade went. Even with the best intentions, a rule set that looks good on the period it was designed on is partly fitted to that period: this is in-sample evidence. The live record starts on the launch date, after the rules were frozen, and nobody can tune them to it. So Stoxa shows the backtest labelled as a backtest, in grey, and never puts it in a leaderboard, a sort or the public table. Since-launch numbers are simulated paper trading at daily closes, after estimated costs, with dividends; a real account will differ by fill prices, timing, taxes and fees. CAGR is shown only after a full year of live days: annualising a few weeks is noise.

When the rules change, the strategy gets a new version with its date, and the record before it stays. You can always see which rules were live on which day.

The risk score

A score from 1 (lowest) to 5, Portfolio123's formula, computed across the house set. For each strategy take the last 252 daily returns of the live track (topped up from the end of the backtest while the track is shorter, and the page says so), and three inputs: the annualised volatility, the maximum drawdown, and the number of holdings. Rank each input across the strategies as a percentile from 0 to 1 so that higher volatility, deeper drawdown and fewer holdings are riskier, average the three, and

where $\bar q$ is the average percentile. The score is relative: it ranks the house strategies against each other, it is not an absolute measure, and a 5 among eight long-only equity strategies is still an equity strategy.

Trading commitment

How much work following the strategy takes. Points for the rebalance frequency (annual 0, quarterly 1, monthly 2, weekly 3) plus points for the average number of trades (buys plus sells) per rebalance, leaving out the first, which fills the portfolio: under 4 trades 0, under 10 trades 1, otherwise 2. A total of 0 to 1 is low, 2 to 3 medium, 4 to 5 high. A monthly strategy averaging 6 trades scores 2 + 1 = 3, medium: about six orders to place twelve times a year.

Liquidity floor and maximum recommended capital

How much money the strategy can absorb before your own orders move the prices. The liquidity floor is the mean 20-day dollar volume of the least liquid 20% of the current holdings (at least one stock). On the assumption, quoted by Portfolio123, that one can trade up to 5% of a stock's daily dollar volume in a day without much impact, and with equal weights:

Following: jump in or scale in

Following a strategy creates a paper portfolio with your capital. Jump in buys every current holding at its model weight at the latest close: you own what the model owns today, including positions it bought months ago at different prices. Scale in starts in cash and only takes positions the model opens from now on; the model's current holdings are excluded. Jumping in is invested at once; scaling in avoids buying names the model may be about to sell, at the cost of holding cash for a while. Both are reasonable; scaling in is gentler for strategies that rebalance often.

Drift and when to rebalance

Your portfolio drifts from the model as prices move and as the model trades. Drift measures the difference, with cash as one more holding:

0 means identical, 1 means nothing in common. The rebalance diff lists the orders that would bring you back: sell what the model no longer holds, buy what it holds and you do not, and trim or add a common position only if its weight is more than 25% (relative) away from the model's. That band keeps you from paying costs to chase small drifts. You accept the whole diff at the latest close or none of it. A sensible routine: accept when the model has traded (you get a notification), and otherwise ignore drift below about 0.1.

Why public holdings are 30 days old

The public strategies table shows the tickers each strategy held 30 days ago, with no weights or share counts; current holdings and trades are for subscribers. The lag keeps the table useful for judging what a strategy owns without turning it into a live trade list.

Judging a strategy that has lagged for a year

Every strategy with an edge spends a lot of time behind its benchmark. The arithmetic is unforgiving. Suppose a strategy truly beats the market by 2 points a year (a large edge) with a tracking error of 8 points (typical for a 20 to 30 stock portfolio). Treating annual excess returns as roughly normal, the chance of a losing year is

and the chance of trailing over five years is $\Phi(-0.25\sqrt{5}) \approx 29\%$. To show the edge with the usual statistical confidence (a t-statistic of 2) takes $(2 \cdot \text{TE}/\alpha)^2 = 64$ years. One bad year says almost nothing about a strategy; one good year says just as little. What does say something: whether the strategy still does what its rules describe (holds cheap stocks, trades as often as expected), whether the style it relies on lagged everywhere (value as a whole had a long drought after 2007), and whether the lag is within what the backtest's drawdowns and rolling windows showed. Abandoning a strategy at the bottom of a normal drawdown is the most expensive mistake in systematic investing.

The claims Stoxa holds itself to

• Leaderboard and public numbers are since launch, and nothing else.
• Backtests are always next to the words "backtest" or "before launch", and are described as in-sample.
• Live numbers are simulated paper trading at daily closes, after estimated costs, with dividends.
• A rule change is a new version with its date; the record before it stays.
• Every rule is public: the definition the simulator runs is served by the API, and the plain-language description is generated from it.