Starting over

In June I rebuilt my allocation engine and wrote about it here. Five months and one more refactor later, I was about to fund it. Before putting money in, I asked for one thing I had never actually looked at in one place: every sleeve of the portfolio, the logic behind it, and the backtest it rests on.

I didn't recognise my own portfolio.

What the overview showed

The factor book was built from eight sleeves: the US market, momentum, size, value, quality, low volatility, and two tactical sleeves for cross-asset trend and macro regime. Laid side by side, the evidence looked like this:

• Only two of the six static sleeves made money over the period I could test them on (late 2017 to mid 2026): the market itself and momentum. Size, value, quality and low volatility all lost money against the S&P 500. They stayed in at 5% each as "factor optionality", which is a polite way of saying I kept them on faith.
• The weights had never been tested as a whole. The final mix was justified by a sweep that used a simpler stand-in for the real sleeves. The study that should have set the weights recommended something else.
• The impressive long-term numbers described a different portfolio. The long-window results came from proxies holding far more equities than the portfolio I was about to fund.
• The versions I could actually buy as a European investor were never fully backtested. Their factor loadings were checked; their returns as a portfolio were not.
• The risk rules existed only in the research. A volatility target and drawdown alerts were written down and never wired into the engine that makes the monthly plan.

None of this was hidden. It was all in the notebooks. It just had never been laid out in one place, and every refactor added more machinery around the same unexamined weights.

The decision

I'm not patching it. The old system stays as a reference: its research, its decision log and its failed experiments are worth keeping. But no weight, rule or default carries over.

Instead I'm building something with a different starting point: a decision engine that has to prove every rule it uses before it may touch money. Every method, whether it decides the long-term mix, a factor tilt or a tactical move, goes through a test lab first. It has to beat the simplest sensible alternative, a global equity fund mixed with a safe asset, on the same data, with the same costs and the same measuring stick. If it can't, it doesn't go in, and the test stays on record.

The other half of the decision: the blog becomes part of the process. Every decision gets a short entry here, including the ones that turn out wrong. This one is the first.

What stays

The accounts stay, and so does the split of roles: a long-term, slow-moving portfolio, and a separate account for tactical allocation. The engine still only proposes; I review and place every order myself.

What changes is what has to be true before anything is bought. The monthly performance reviews I announced start with the engine's first live month, whenever the engine has earned it; that might be December, or it might not.

The next entry lays out the design and the rules it has to live by.