Hi YXI friends,

A month ago, I started publishing Multimodel Signals — four independent models voting on each name, every morning, before the US open. Since then, I have been working on a Systematic Portfolio based on these Multimodel Signals.

Today, I am very excited to announce that our Systematic Portfolio is now live. This is a unique, proprietary approach you will not see elsewhere. We already posted the portfolios in our previous email, but here is an in-depth note on how the portfolios are constructed, their backtest results, and what you can expect to see over time.

DISCLAIMER: This newsletter is intended for educational purposes only. Any information or analysis in this note does not constitute an offer to sell or a solicitation of an offer to buy any securities. Nothing in this note is intended to be investment advice, nor should it be relied upon to make investment decisions. Any opinions, analyses, or probabilities expressed in this note are those of the author as of the note's date of publication and are subject to change without notice.

If you read the June introduction, skip this.

Four models look at every name, and they are built differently on purpose:

Every day I publish how many of the four agree with the published direction. Three or four agreeing is Broad. A narrower split is Mixed.

None of the four is the whole truth. Together, they get a lot closer.

One month in, the thing I would add to that letter is this: the disagreements have turned out to be more informative than I expected. A name where all four agree behaves differently from a name where two do — not just in hit rate, but in how it behaves when it is wrong. That observation is what the portfolio is built on.

Systematic equity portfolios are typically built in one step, defined by their factor (e.g. momentum, value, quality).

Ours is built in two.

Step one is our Multi-model Signals, which assess each asset on its own merits, with four models, and produce a published record: is this name long or flat today, and how broadly do the models agree? That record exists whether or not any portfolio is ever built from it.

Step two builds a portfolio only from names the record already calls long, and only where at least three of the four models agree. It never starts from a raw price screen.

Why the extra step?

Because momentum, or any single price characteristic, tells you what a price has done. It cannot tell you whether the move is supported.

A name that has run 40% because a trend is genuinely in place (our Trend Model) and the regime supports it (our Market Regime model), and a name that has run 40% into a hostile regime on one model's say-so, may look identical on a price screen. But they are not identical, and four models arguing about it is a better way to tell them apart than a single sort.

Moreover, the two steps fail independently. If the portfolio construction is wrong, the signals are still right, and you can see it. Mistakes stay hidden in a one-step process.

There are five stages.

1. The record must call it long.

The portfolio can never hold something the signals page shows as flat.

2. At least three of the four models must agree.

A Broad model consensus call gets through; a Mixed one does not.

3. Rank the qualifying candidates.

We rank the candidates by how many models agree (i.e. we prefer unanimous consensus vs 3/4), with 63-day momentum as the tiebreak.

Up to 10 names are held in each book.

For the Macro & Megacap portfolio, we enable an additional correlation-aware selection, thanks to its large asset universe. New entrants are picked by rank minus their correlation to what's already held. This does not apply to the Commodities and FIDA portfolios, as they are inherently single-themed and have a smaller universe.

4. Weight by risk, not by cash.

Each holding is sized so it contributes an equal share of the portfolio's risk.

A volatile name gets a smaller slice than a quiet one, so that no single position dominates what happens to the portfolio day-to-day.

On top of that sits a cap on any single position, applied each time the book is re-optimised. The exact cap depends on the number of holdings. At ten holdings, no target position exceeds 15% of the invested book (not including cash), and no target ever passes 25% of the invested book, however few names are held.

5. Scale the whole book to a volatility aim.

Every book runs at less than 100% invested, deliberately. At launch: Macro & Megacaps 79% invested, Commodities 78%, FIDA 50%.

The book aims for a long-term volatility level at or below that of the S&P 500. That cash is the volatility aim doing its job: the whole book is scaled so that its risk lands near a target. When the names it holds are jumpy, it holds less of them. When they are calm, it holds more.

We use the rules below to reduce portfolio turnover, with a small trade-off in returns. However, I must warn that the systematic portfolio approach is still very active.

1. Falling Ranking

A holding is droppable once its rank falls to 15th or worse.

When slots open, at most two new names enter per re-optimisation (on the first trading day of the week).

This costs a little return but saves a lot of turnover.

2. Multi-model Signal Changing

A position is sold the same day, any day of the week, if the published record turns it flat, or if fewer than three of the four models still agree with it.

3. Rebalancing: Small moves are not traded

If a holding's target weight has moved less than 5 percentage points (e.g. from 4.4% of the book to 9%), it is left where it is. That is a wide band on purpose to reduce turnover costs at very little impact on returns.

The book re-optimises once a week, on the first trading session of the week (i.e. Monday or Tuesday after a holiday). That is when names can enter, weights are re-solved, and the cash level is reset.

But the exit check runs every day. A holding is sold the same day, any day of the week, if either:

The second rule means a name can be sold out of the book while the signals page still shows it long, because agreement has thinned below the threshold.

The proceeds sit in cash until the next re-optimisation. When a name exits, the proceeds stay in cash. Nothing new is bought before the weekly decision.

What you get, and when. We publish our Systematic Portfolio with the Multi-model Signals in the same note every trading day before the US stock market open. Every Monday, I will share both the portfolio P&L and the per-ticker multi-model signal performance.

Over the full sample period, the book compounded at 32.5% a year against SPY's 15.3%. At the same time, the systematic portfolio experienced volatility of 14.6% versus SPY's 17.2% and a maximum drawdown of −16.8% versus SPY's −32.0%.

This was achieved while sitting in cash 32% of the time on average. However, this is a highly active book, once you account for the weekly portfolio optimisation rebalancing and intra-week exits. Assuming a 5bp trading cost, the backtest suggests a 0.9% annual trading cost (already deducted from CAGR).

Two things to keep in perspective. A Sharpe of 2.01 is a very high number, and you should treat it as what it is — a backtest, not a promise. The book held an average of 8.4 names, not ten: the agreement gate frequently does not pass ten candidates.

Its worst calendar year was 2022 at −12.80%, when the index fell 18.67%.

16.7% a year against SPY's 15.3% is an edge of roughly 1.4 points, but there are important details I must flag. Trailing windows show a CAGR of +11.6 points outperformance versus SPY over one year, but −17.6 points over three (underperformance) and −1.6 points over five (small underperformance).

What it actually delivers is the risk column: 14.1% volatility against 17.2%, and a −15.3% worst drawdown against −32.0%, with a quarter of the book in cash on average.

If you look closely at the chart, Commodities often trade in the opposite direction to the S&P 500. This is an important diversification benefit, as in 2022 it returned +19.6% while the index fell 18.7%.

That is the case for owning it, i.e. not that it beats the market, but that it does something different when the market is falling.

2016-2026

29.4% a year against 15.3%, with a −25.9% maximum drawdown against SPY's −32.0%.

Because we have such a small universe at the moment, the backtest average holdings are only 2.6 each year with an average cash 53.2%. The latter is more of a volatility targetting outcome.

2021-2026

As Bitcoin & Ether rose exponentially in their early days, I want to provide a 5-year view to set more realistic expectations.

This is the more honest window, since Crypto became more institutionally traded during this period.

17.3% a year against SPY's 12.7%. Sharpe falls from 1.58 to 1.00, with a drawdown similar to SPY's.

While the Commodities and FIDA portfolios appear “inferior” to the Macro & Megacaps, below shows an important function they help serve.

As we run three Systematic Portfolios (Macro & Megacaps, Commodities, and FIDA), allocating money to two or all three portfolios can unlock significant diversification benefits.

In our backtest simulation, we find that adding the Commodities portfolio can improve the maximum drawdown significantly compared to the standalone portfolio. MM+CO and CO+FIDA are the only combinations with no losing calendar year in the sample.

At the same time, adding the FIDA portfolio can substantially boost returns. MM+FIDA returns the most within the combinations and charges you a −11.29% year for it. That’s a similar CAGR to MM alone, but with better maximum drawdown and worst-year performance.

Combining all three books (equal allocation and annual rebalancing between the books) yields the highest Sharpe ratio and the lowest volatility.

On the rebalancing note - my backtests show that quarterly rebalancing vs annual rebalancing does not really create a meaningful difference in Sharpe, but I personally would not let the books drift beyond the annual rebalancing. One book could easily dwarf another over time due to outperformance.

While our volatility target aims to undershoot that of the S&P 500 (while delivering better returns and drawdowns), one may have a higher or more conservative risk tolerance in their vol preference.

Here is how the backtest results would look if the positions were scaled to end up with more or less cash.

Note that the “average cash (as shipped)” is the average cash holding in the backtest period using our approach, not our launch portfolio cash level announced today.

By increasing risk-asset holdings, one typically gets higher returns at a similar Sharpe ratio, but the trade-off is higher volatility and a worse maximum drawdown.

It is long-only and unlevered. No shorts or hedges. In a broad bear market the only defence is cash.

The 2× tier on the signals pages does not apply here. Every portfolio holding is a 1× position with an individual position cap.

Concentration is real. Selecting up to 10 names is much more common than one may be used to seeing. For the FIDA portfolio, it is often holding only 2-4 names.

The turnover is not tax-friendly. Portfolio turnovers in a taxable account, in a country that taxes short-term gains, can result in a lot of tax costs.

Midweek exits require the ability to act. If one can only trade on a certain day of the week, their results will differ from mine.

A backtest is a backtest. The next 10 years may not repeat the bull market of the past 10 years. Backtested results do not guarantee future performance.

I might just be wrong. Four models agreeing is a better bet than one. It is not a guarantee, and there will be months where all four are wrong together, because they are all reading the same market.

Please help me improve the service with your immediate feedback - thank you.