This page exists so you can check our work before you pay for it. The cost assumptions are stated. The results are given as a range rather than a best case, and the one crash this model failed to protect against is described in as much detail as the ones it handled. The rules themselves stay ours.
Four US-listed ETFs, equally weighted at 25% each, reviewed on a fixed schedule of roughly one review a month. When the portfolio changes you receive the new holdings and the trades required to get there, through a private Telegram channel.
Between reviews there is nothing to do and nothing is sent. Long stretches of silence are the normal state of this model, not a gap in the service.
The model may only ever hold names from one fixed list of thirty-two, and that list does not change. Every one of them was already trading before the backtest begins — which is why survivorship bias, the flaw that quietly inflates most published backtests, cannot occur here.
The composition of the universe is part of the model and is not published in full. Subscribers see the holdings the moment the model selects them.
The model ranks the universe on a fixed set of quantitative criteria and holds the four it selects, equally weighted. There is no discretion at any point: no position is added, removed or resized because of a view, a headline or a conversation.
A regime filter sits above the selection. In adverse conditions the portfolio moves entirely to cash and stays there until a later review says otherwise, which is why long stretches holding nothing are a normal state of this model rather than a fault in it.
Equal weights. Long only. No leverage. Four positions, never more.
The criteria themselves, their settings and the full universe are not published. They are the part of this work that took years and that a subscriber is paying to have applied, and publishing them would hand the product to anyone with a spreadsheet. What is published instead is everything needed to judge the result: the assumptions, the costs, the full history, and the distribution of outcomes across the whole neighbourhood of settings we tested.
The settings were taken from peer-reviewed finance literature published decades ago and frozen before the strategy was ever tested, rather than searched for because they performed well on this data. They have not been changed since. We re-optimised them three times over the course of this project, and out-of-sample results were worse every time, which is the ordinary fate of parameters fitted after the fact.
Costs of 5 basis points per side on every trade.
Cash earns the real 3-month USD rate whenever the model is out of the market. Omitting this would understate a strategy that spends time in cash — and most published backtests omit it.
Signals are computed on closing prices and acted on the following session.
Period: 28 September 2007 to 18 September 2026 — the full history available for this universe, 2008 included.
These are backtested results. They are not an executed track record: no orders were placed and no fills occurred.
| Strategy | S&P 500 ETF price only | |
|---|---|---|
| 2008 calendar year | +3.0% | −38.4% |
| Full 2007–2009 crisis | +4.4% | −55.4% |
| 2022 bear market, Jan–Oct | +22.1% | −24.9% |
| Covid crash, Feb–Mar 2020 | −27.9% | −33.8% |
Through 2008 the strategy never traded below its starting value. The index was down 47.9% at its worst point that year.
The 2020 crash is the one this model did not protect against, and the reason is structural rather than accidental. The portfolio is reviewed on a fixed schedule of roughly one review a month, so a collapse that completes inside four weeks finds it fully invested with no opportunity to react.
It lost 27.9% against the index's 33.8% — protection, but not much of it. It finished that year up 23.4%.
| Strategy | S&P 500 ETF price only | |
|---|---|---|
| Annualised return | 13.8% | 8.8% |
| Sharpe ratio | 0.85 | 0.53 |
| Worst drawdown | −28.6% | −56.5% |
| Annualised volatility | 17.0% | 19.8% |
| Years ahead of the index | 11 / 20 | — |
Both columns exclude dividends. The strategy is not credited with the dividends of the ETFs it holds, and the S&P 500 ETF is shown on price alone. With dividends reinvested the index would gain roughly 2 percentage points a year; the strategy would gain from the dividends of its own holdings while invested.
A single set of settings can flatter any strategy. We tested 54 neighbouring sets, varying every setting the model has, and the frozen set sits at the 91st percentile of that distribution.
| Frozen set | Median of 54 | Worst | |
|---|---|---|---|
| Sharpe ratio | 0.85 | 0.63 | 0.23 |
| Worst drawdown | −28.6% | −35% | −51% |
Treat the median as the honest expectation. There is no reason the future should draw the same percentile as the past, and a service that quotes only its best configuration is telling you about its luck rather than its method.
APA Quant is based in Milan. We build systematic models and run them on proprietary capital before they are offered to anyone, and their settings are frozen before release. Our research is led by a portfolio manager with a degree from Bocconi University and more than 21 years of experience in portfolio management.
That background explains the method, not the outcome. Nothing here is discretionary: the models decide, and what they decide is published as it stands — including the months they get it wrong. The cost assumptions, the record and the limitations are all on this page precisely so that you can judge the system rather than the person behind it.
These are general signals, identical for every subscriber. They are not personalised investment advice, they are not tailored to your circumstances, risk tolerance or objectives, and we do not know what else you hold.
We make no forecasts and promise no returns. This strategy is not built to beat the index in a rising market — it did so in 11 of 20 years — but to lose substantially less when the market falls. Judge it over a full cycle, not a quarter.
Past performance and backtested results do not guarantee future results. You can lose money, including on positions shown in the channel.