All three published strategies are ahead of the market, and the Núcleo holds up across all 21 start dates

As of 9 October, the three strategies we publish are ahead of the house benchmark. Since seeding on 18 June, Núcleo S&P 500 is at +5.43%, Cesta Diversificada at +5.03% and Defensivo Táctico at +3.06%. The equal-weighted market of the panel, 606 names measured over the same dates, returned +2.73%. All of it in simulation and net of turnover costs.

The change of the week belongs to the Núcleo. In early September we reported here that it had crossed below the benchmark. It has crossed again, this time upwards: it rose 3.20% over the last week and now stands 2.70 percentage points ahead. The Cesta is 2.30 points ahead of the market and the Defensivo 0.33. Five weeks ago none of the three was above it.

They got there without drama. The worst drawdown since seeding is 6.4% for the Núcleo, 2.7% for the Cesta and 2.3% for the Defensivo.

The second piece of news is a test we owed the Núcleo. A monthly backtest hides 21 backtests: depending on which day of the month you start the calendar, the same strategy buys on different dates and the result changes. So we repeated it on all 21 possible start dates, from January 2022 to August 2026, net of costs. It beats the market on all 21. The median is +16.0% a year against +8.4% for the market, and the worst start date gives +14.2%. The Sharpe ratio, measured on daily values, has a median of 1.06, and the maximum daily drawdown a median of 15.2%. The edge over the market is statistically significant on 15 of the 21.

With that test we have also settled on a single figure for the Núcleo, and we picked the most demanding one. The backtest shown on the site now measures the Sharpe ratio and the maximum drawdown on the portfolio's daily value, which is what someone holding it would feel, rather than on one point a month. Since 2020: +18.0% a year against +12.2% for the market, a Sharpe of 0.93 and a maximum drawdown of 36.7%. On monthly blocks the Sharpe came out at 0.98; we are keeping the daily one.

We have tidied the catalogue. Two experiments are retired: Crecimiento agresivo, which stood at −6.2% and never passed validation, and Núcleo apalancado, which moved exactly like the Núcleo and could not teach us anything new. They do not disappear: they stay in the lab with their track record and a “Retired” label, because what we kill we do not hide. Defensivo + carry, which turned out to be the same portfolio as Defensivo Táctico under another name, moves to the lab as an experiment. Three published strategies remain, and each one is a different thing.

In crypto we learned something useful. The momentum strategy reached +53% in September and gave a good part of it back, so we tested five ways of cutting drawdowns on six years of data, including the coins that disappeared. Selling each coin when it falls, which is what instinct asks for, does not help. What does improve things, in all 14 calendar variants we tried, is giving less weight to the most volatile coins and exiting the same day Bitcoin loses its trend: in the backtest the Sharpe goes from 0.88 to 1.05 and the worst fortnight from −49% to −35%. It is now running in parallel as a new experiment in the lab.

And there is a new experiment with language models. In the AI tournament the models can take shelter in a 50% probability. Here they cannot: a model follows the Núcleo portfolio and every morning, before the open, has to rank its stocks from the one that will do best to the one that will do worst. It is scored against the whole basket and against three mechanical rules with no AI. It started on 12 October and is in its registration phase.

What we are not saying. Less than four months in simulation proves nothing: with so small a sample, none of these live edges can yet be told apart from chance. And the 21-start test begins in 2022 on purpose: 2020 and 2021 were bad years for the Núcleo's signal, and with them included the result is less conclusive. None of this is investment advice. All strategies remain in simulation, with no real money behind them, and the backtest figures are historical. The numbers are at soberquant.com/bots and the experiments at soberquant.com/lab.