Two patterns picked companies that beat the simple alternative in both halves of the test and in most individual years. Here is what they buy today.
What the numbers below mean
There is a pool of companies you could buy: US shares above $3, trading more than $1m a day, worth between $100m and $20bn. About 1,400 of them at any moment. Buying all 1,400 in equal amounts is the simple alternative, and it earned a year. Each of these two patterns instead ranks the pool and buys only the best fifth, roughly 280 companies, rebuilt every three months. Holding half your money in each pattern earned a year. The difference, a year, is what the patterns were worth.
Tested on companies over years, including ones that have since delisted or gone bust, because dropping the failures is how a test flatters itself. Dealing costs are already deducted.
The buy list
The best fifth of the pool on each pattern, as at the last quarter end. Not advice, and not a portfolio: a starting list to argue about. Every name is a real company you can look up.
Growth of £1
What £1 became. Rebalanced quarterly with dealing costs deducted, against buying all 1,400 companies in equal amounts over exactly the same dates.
Does the ranking work all the way down?
Split the pool into fifths by the pattern. If the pattern is real, the second fifth should beat the third, and so on to the bottom. A signal where only the top bucket works is usually a fluke.
Year by year
Percentage points by which the best fifth beat buying everything, in each year. This is the test of whether you could actually live with it.
The numbers
In sample is 2016 to 2021, out of sample is 2022 onwards. Nothing was tuned on the second period. The t-statistic is on quarterly excess returns over the universe; above 2 is the usual bar, and the blend clears it most comfortably because the two signals are negatively correlated in their excess returns.
What did not survive
The Arena
Two books of £1,000,000. Same prices, same costs, and the signal portfolios alongside as the thing to beat. Every position carries a thesis, because a position without a reason teaches you nothing when it closes.
Open a position
Marked at the last close in the snapshot and charged 10 basis points each way. Fictional money.
How this was tested
Every quarter, score every company in the universe on the signal. Buy the top fifth, equally weighted. Hold three months. Rebalance and charge dealing costs on whatever changed. Compare against holding the whole universe equally weighted over exactly the same dates.
The universe is US-listed common stock above $3, trading more than $1m a day, with a market value between $100m and $20bn, rebuilt every quarter from share counts reported on SEC cover pages multiplied by the adjusted close. Membership is dated, so a company only appears once the market could have known the numbers. Delisted companies stay in the history, which stops the test flattering itself by quietly dropping failures.
Fourteen candidate patterns were tested, all drawn from published research rather than invented here, because inventing a signal and testing it on the same data is how you fit noise. Four were promising enough to examine closely. Two survived.
Three bugs found while doing this
The position sizer could switch itself off permanently. A losing run drove the Kelly fraction to zero, and a sizer at zero takes no trades, so it never receives the results it needs to revise itself. It had silently stopped trading in 2018 and produced a flattering result from the years before. Zero is now floored.
The alpha calculation compared each quarter's returns against the previous quarter's factors. That removes no factor exposure at all, and reported the average return as if it were skill. Every alpha figure in the first version was wrong and far too high.
Momentum was measured in rebalance periods rather than months. On a quarterly rebalance, twelve-month momentum was quietly measuring three years.
All three are fixed and the figures here are from the corrected code. They are worth stating because you should trust a number more when you know what was checked, and I would rather show you the errors than the polish.
What would change the answer
Thirty-eight quarters is a small sample. A t-statistic near three over that span is meaningful but it is not proof, and both of these patterns are well documented, which means their returns may compress as more money chases them.
The momentum list in particular buys small companies after very large rises. Dealing costs there are charged at 30 basis points, which may be optimistic for the smallest names. If you run this, the first thing worth measuring is what you actually pay to get in and out.
Built on free data. SEC EDGAR bulk filings and XBRL company facts, Alpaca market data on the free tier, and the Ken French factor library. No paid data and no subscription.
Not advice. Every figure here is a historical simulation. Past patterns are not a promise, no strategy shown will necessarily make money, and the Arena is a game played with fictional money. Speak to a regulated adviser before risking real capital.