The design: an engine that has to earn its rules

The last entry explained why I threw out my allocation engine. This one lays out what replaces it, and why it's built the way it is.

One question first: what counts as good enough?

Everything starts with a written policy, the way institutional investors start with an investment policy statement. Mine has one rule at its centre:

Average drawdowns should not be larger than a year's return.

In numbers: the compound annual return, divided by the average of each year's worst peak-to-trough fall, should be at least 1. This measure has a name, the Sterling ratio, and it captures what I actually care about: not how volatile a portfolio is on paper, but how deep the holes are that I have to sit through, compared with what I get for it.

The rule has teeth. A broad equity index has historically scored around 0.6 on it: roughly 8 to 9% a year, against an average drop of around 14% within the year. So the rule forces the design to do something. Hold fewer equities, or add parts that cut drawdowns more than they cut returns. Which works better is exactly what the engine has to find out.

The worst crash in a decade isn't part of the rule. It can't be predicted, so it's reported for every option but doesn't decide anything. And if loosening the rule to "drawdowns up to one and a half years of return" buys a much better portfolio, the engine will show it, and I can change my mind on evidence.

The gatekeeper: a test lab

The heart of the design isn't the part that picks investments. It's the part that decides whether a way of picking investments is allowed in at all.

Backtests lie easily. Try enough variations and one will look brilliant by chance. Bailey and López de Prado call this backtest overfitting, and they built tools to measure it. The lab uses them, along with a few rules practitioners have learned the hard way:

• Fixed rules before testing. Write down the method first, then test it.
• Out of sample, always. Decisions only use data available at the time, and a final stretch of history is tested exactly once, at the very end.
• Plateaus, not peaks. A method qualifies only if its neighbours work too, not just one lucky setting.
• Costs and taxes included. German taxes on every sale, fees, currency conversion.
• One yardstick. A global equity fund plus a safe asset, mixed to just meet the drawdown rule. Anything more complicated has to beat that, or it doesn't get in.

What the lab may admit

Three kinds of decisions, each with its own evidence base:

1. The long-term mix. The starting point is the global market portfolio: every asset class at its market weight, the allocation that needs no forecast. Simple alternatives such as equal weights or risk parity get tested against it. One popular method is ruled out from the start: optimising on estimated returns, which loses to plain equal weights out of sample (DeMiguel, Garlappi and Uppal, 2009).
2. Factor tilts. Value, momentum, quality, small caps, low volatility. Each has decades of evidence, and each shrank after it was published: McLean and Pontiff found anomaly returns about 58% lower after publication. A tilt gets in only if it still clears a high statistical bar after that haircut, and only if it makes the whole portfolio better, not just itself.
3. Tactical moves. Trend-following across asset classes has the strongest evidence of any tactical approach, a century of it (Hurst, Ooi and Pedersen, 2017). Its value shows up mainly in long, grinding crashes, and it costs something in bull markets. How large this part should be isn't my gut call: the lab tests every size from zero to half the portfolio and takes the smallest one that clearly helps.

What it will never do

Send an order. The engine proposes a basket each month; I review it and place it myself. It also stays out of the parts I don't control, like a robo-advisor portfolio. Those show up in the monthly reports, but not in the engine's targets.

How it gets built

In five steps, each ending with something I can check:

1. data, the test lab and the yardstick
2. the long-term mix and the factor tilts
3. the tactical part and the risk checks
4. rebalancing, taxes, the connections to my brokers, and reporting
5. the first live month

Step one starts now. The next entries go into the details: how the lab measures, and the papers each rule rests on.