Hi YXI friends,

At the end of June, I retired the single-model machine-learning signals and moved everyone across to Multi-model Signals.

Before I close that chapter for good, I want to put a scorecard on the record, as both a review and a case study for understanding how AI/ML models work. This is the story of what I was trying to build, where it delivered, where it let me down, and what those lessons are pushing me to build next. After all, the ML model is still a key component in our Muti-model Signals.

After this review, I will write a separate in-depth piece on the MS approach.

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.

Here is the headline first, because everything else is shading on it.

Over the full period and the whole fleet, the model's tracked positions returned +20.3%, against +18.4% for simply buying and holding the same names — and it did so at a smaller drawdown (−13.8% vs −16.1%) and a far better risk-adjusted return (Sharpe 0.93 vs 0.66).

It beat the benchmark on return and on risk. But the shape of that win matters more than the number, and I'll come back to it.

I did not build a model to beat a raging bull market. I built a model to survive a full market cycle, i.e. to be Risk-On when the weight of evidence favours the upside, and to step aside when it doesn't, so that the bad quarters do less damage.

The right way to judge a system like that is across an entire cycle, not across any single week. Some weeks it will look too cautious. That is the cost of the insurance, and the only fair question is whether the insurance paid for itself over time.

The Verdict: On the full-cycle test, it did. On the individual episodes, it was much more mixed.

Let's go through them.

We launched our Macro & Megacap tickers between April and May 2025 (with late additions to the equity and sector indices after the New Year).

You'll notice the signals change shape across three phases:

When most of the market was bearish in the March-April 2025 crash, the model turned Risk-On decisively in April and May 2025.

It held that call through most of October, and it was right. The indices didn't crater; instead, they reached a new high in the New Year.

It then went Risk-Off in February 2026, correctly, as the market sold into the US–Iran conflict. SOXX turned bullish on April 9th, right after the ceasefire, in time for a historic run, and QQQ followed on April 24th.

The lowlight is the one that stung: SPY sat out most of the April V-recovery.

Two things caused it. First, the model read the initial rally as a sell-the-rally regime, because oil stayed stubbornly high and, through the correction, higher oil, yields, and the dollar had been driving risk assets down. Therefore, the model kept fading the bounce.

Second, the smoothing that prevents the signal from whipsawing means the earlier bearish reading still carries weight, and the model waits for stronger evidence before committing to a new position. Together, that delayed the long by about two weeks — a lag that cost real upside.

SPY - Price top pane, model position bottom pane

SOXX

QQQ

New MS Model Backtest Walk-forward comparison

Here, we are simulating the MS results as if they were a live environment.

The new Multi-model Signals approach we are using now shows a much more stable signal setup, with fewer positional flips and better trend capture.

However, some false signals are inevitable. From the above chart, you can see how well an asset responds to our approach - some more effectively than others.

The way we have done our robustness backtests is to ensure our approach can work in all weathers - meaning it doesn’t just fit well in a raging bull market, but also prepares for the eventualities of sideways or bear markets.

The crypto complex tells a different, and much better, story. This is where the model earned its keep.

As the whole complex rolled over, the model's tracked positions returned −1.9%, while buy-and-hold lost −31.8%.

It side-stepped roughly thirty points of losses. Underneath the top line:

If you want one piece of evidence that the model does its core job of protecting capital when the tide turns, this is it.

The frustration in this period was that the Bitcoin bearish signals were busy during the initial downturn. It flipped in and out through the October–November consolidation before I'd deployed the smoothing upgrade.

BTCUSD

ETHUSD

COIN

We launched COIN’s ML model near the top of its cycle. It was the worst name in the book at −49%, though even that beat the −64% you'd have taken by simply holding.

COIN taught me a rule I now weigh heavily: it is hard to model a name with fewer than five years of history, because the model has too few analogues to learn from. We are keeping COIN in this service only for continuity.

New MS Model Backtest Walk-forward for Crypto

Under the MS model, one would have sat almost entirely out of BTC, ETH, SOL, XRP and COIN in November 2025, with very few attempts to play the bounce.

HOOD and MSTR appear different because they currently use single-model strategies. This is because some single-model strategies were tested to be more effective than the multi-model for certain names. We will continue evaluating the results in a live environment.

I have to be careful and honest about commodities, because it's where the lived experience felt worst and where the numbers are the most misread.

On many names, the model beat or defended buy-and-hold:

The entire difficulty sat in three names: GLD, SLV and SIL — the precious-metals ETFs that went parabolic into the New Year and then topped violently.

So why those three, so badly?

Because a parabolic top is the one regime where a trend-and-mean-reversion model faces a genuinely impossible choice. Stay long to ride the parabola, and you take the crash on the way back down, which is what happened on GLD and SLV.

Exit early to dodge the crash, and you miss the biggest, final leg of the run. The move that captures a near-vertical rise and the move that dodges a two-day 30% collapse are the same move. When GLD dropped 15% in two days and silver 30%, nothing still riding the trend could have avoided it, and anything fast enough to avoid it would have bailed on the run weeks earlier.

Here is the tension I live with as the person building this, and as an investor myself. I want two things that cannot both be true at the same time:

Those are opposite instructions.

Tune for fast protection, and you exit on every scary dip, missing the rebounds. Tune for participation, and you hold through the one dip that doesn't come back. You can optimise for protection or for participation. But a single signal cannot maximise both, and the sharper the parabola, the more brutal that trade-off becomes.

New MS Model Backtest Walk-forward for Commodities

I want to be completely clear about what the new model can and can't do here. I had tested this exhaustively before the upgrade, with every protection idea I could build (e.g. regime gates, trend gates, drawdown stops, volatility-based sizing) across the whole fleet and multiple market cycles.

The finding was unambiguous. No rule reliably protected these parabolic-top drawdowns without giving up more in ordinary uptrends than it ever saved in the crash. The protection that looks brilliant with hindsight only works because it already knows the crash is coming (i.e., overfitting).

Therefore, to set the right expectations, I won't tell you Multi-model Signals fixes the precious-metals top.

What it does do is flip far less in the chop after a top, blend four independent engines so no single model's blind spot dominates the call, and capture the cases where a regime read genuinely helps.

The comfort I take, and the fair conclusion, is this: the commodities pain was not a model that was poor by design. It was a specific, unusually hostile regime, ie., a parabolic run and violent top in three correlated names. It is not systematically solvable within our means.

The model's value is defensive, and it showed up where you'd want it: in the markets that fell apart.

Crypto is the clearest case (−1.9% vs −31.8%), but the pattern repeats across the fleet: META −10.7% vs −16.8%, and PLTR is actually positive at +4.6% while holding lost −6.5%.

In commodities, the miners carried it: GDX +21.9% vs +3.1%, PPLT +38.9% vs +3.4%.

The cost of that insurance is participation in the strongest, lowest-drawdown runs. Macro & Megacaps lagged its benchmark in absolute returns, +21.9% vs +30.2%, though it enjoyed a smaller drawdown (-10.7% vs -15.5%) and a comparable Sharpe (1.1 vs 1.2).

And there were a few genuine misses I won't dress up: AMZN −6.5% vs +22.4%, IGV −9.1% vs +4.9%, and HOOD −23.3% vs −3.4%.

This is an all-weather, risk-managed model. It wins by losing less when things go wrong, and it pays for that by capturing less when things go right.

Netted over the full cycle, it came out ahead on both return and risk, which is exactly the trade a risk manager is supposed to make.

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