When you have spent months tuning a system's parameters, a frightening question shows up: am I optimising something that is structurally wrong?
There is only one way to answer it. Build the opposite and measure it with the same rigour.
The experiment
Cerberus enters against the move, averages down with a multi-position grid, and its profit per cycle is capped while the loss is not.
Aegis inverts it on every axis:
| Cerberus | Aegis | |
|---|---|---|
| Direction | Against the move | With it |
| Positions | Grid, several | One only |
| Size | Fixed lot | Risk-derived |
| Exit | Shared target | Server-side stop |
| Averages down | Yes | No |
It is not a variant. It is the negation of the design.
What it measured
Real M1 gold, 2019 to 2026 — 3.4 M candles, 1,026 trades:
| Cerberus (live, 904 trades) | Aegis (backtest, 1,026 trades) | |
|---|---|---|
| Win rate | 60.1 % | 40.6 % |
| Average win | +$1.05 | +$33.25 |
| Average loss | −$6.74 | −$21.98 |
| Win/loss ratio | 0.16 | 1.51 |
| Expectancy | −$2.06 | +$0.44 |
| Drawdown | 52.6 % | 23.8 % |
The architecture change did exactly what it was supposed to. The asymmetry inverts: average win becomes larger than average loss, expectancy per trade changes sign, and drawdown drops to less than half. At under half the win rate.
It is the clean demonstration that win rate is not the variable. Aegis is right 40 % of the time and makes money; Cerberus is right 60 % and loses.
And we still did not deploy it
A profit factor of 1.04 over seven and a half years is a coin flip with commissions.
Worse: three of the six periods lose money. And every configuration swept loses money between 2019 and 2023. The entire positive result comes from 2024–2026 — the same recent gold regime in which every system in this project looks good.
That last part taught us the most. When two systems of opposite architecture both look good in the same time window, what you are measuring is probably neither of them: it is the window.
Aegis is not deployable on real money. We documented it anyway.
Why a negative result is worth publishing
A negative result measured properly is worth more than a positive one measured badly, for a practical reason: it bounds the search. We now know Cerberus's problem is not "fade versus follow". It is averaging with a grid, and no amount of direction tuning fixes that.
That conclusion can only be bought by building the whole experiment and accepting whatever comes out. Had Aegis worked, we would have a new product. It came out mediocre, and we got something almost as useful: knowing where not to keep looking.
The alternative — spending another six months tuning Cerberus parameters — would have been more comfortable and would have answered nothing.
From building Cerberus at Tuurt Labs. High-risk instrument: trading leveraged CFDs can cost you your entire capital. This is not investment advice.