How Conduit Was Built: One Trade, One Rule, Two Decades of Data

Conduit is the first algorithm Kingdom Portfolios runs on its own accounts, and it is the clearest example we have of the doctrine in Built, Bought, or Blackboxed. It was built, in the specific sense we mean by that word. Here is how, told as method, because the method is the transferable part.

It started as a hand-traded rule

Long before there was code, there was a habit, and it runs against most people's instinct. When a market is already leaning one way on the higher timeframe, and the shorter timeframe then stretches hard in that same direction, the ordinary reading is "too far, fade it." The founder's reading was the opposite: a stretch that agrees with the bias above it is usually the move continuing, not ending. So he would go *with* it rather than against it, and he would make the market prove it first by placing the order above the stretch instead of at it, so that a push which stalled and rolled over never became a trade at all. Small fixed stop, a target that paid a little more than the risk, one trade at a time, no second position until the first resolved.

That is the whole idea and it fits in a sentence, which is the first test any idea has to pass before we will build it: align with the higher timeframe, enter on continuation through a stretch, and let price confirm before you are in. It is a continuation model, not a reversal one, and the difference matters because the two need opposite things from the market.

Then it became mechanical

The hand-traded version had judgment in it: "already leaning", "stretches hard", "prove it". Making it mechanical meant replacing each of those with something a machine could measure without asking.

For the bias, a higher-timeframe oscillator reading that has to agree with the direction of the trade before anything else is allowed to happen. For the stretch, the same oscillator on the trading timeframe, pushing into a defined zone rather than merely crossing a midline; comparing the two readings is the whole multi-timeframe part, because a stretch only counts when the timeframe above it is pointing the same way. For "prove it", a resting order a fixed fraction of an average range beyond the signal bar, alive for a set number of bars and then cancelled. That last piece is the quiet one. It means an excursion that stalls and rolls over is never a trade, which is a price-based second opinion sitting on top of the indicator.

Then a defined distance for the stop and a defined ladder for the target. The point was never to make the rule cleverer. It was to make it *reproducible*, so the same rule fires on the same bar every time and the results can be trusted to mean something. It also has an honest cost the founder named before we tested anything: entering into strength means a deeper adverse excursion on the pushes that fail, so the measure that matters for this logic is how often it reaches its floor, not how much it makes when it is right.

Then the AI built it

With the rule written as rules, the code was largely written by AI, under direction. It is the part of the story people expect to be the hard part and it was the easiest. Writing a strategy to a spec is exactly what a modern coding model is good at. What it cannot do, and what we did not let it do, is change the spec. Every time the harness surfaced something surprising, the question went back to a human: is this what the rule means, or has the code drifted from the rule? That is the working relationship we mean by human-led and AI-built.

Then it was split into legs

One rule is not one strategy. The same doctrine behaves differently on gold than it does on an index or a currency pair, because the instruments breathe at different speeds, carry different costs and trend for different lengths of time. So Conduit is not run as a single thing. It is run as legs: the same rule, expressed once per instrument, each one measured on its own history as if it were the only strategy we had.

The first instinct is the obvious one: tune each leg until it is excellent. Find the best stop for gold, the best expiry for the index, the best zone for each pair, and stack the winners. We did exactly that, carefully, and then we held back a stretch of data the tuning had never seen and checked whether the tuned settings actually did better on it.

They did worse. Not slightly, and not randomly: in our own walk-forward the ranking of tuned settings *inverted* on the unseen data, so the setting that looked best in the tuning window was reliably one of the poorer choices afterwards. A tuned parameter had become a description of the past rather than a property of the market. The untuned default held up better than the thing we had worked hard to improve. That result changed how we build every leg: we now treat a per-leg parameter as something to justify, not something to optimise, and a leg that only works at one precise setting is treated as evidence against the leg rather than in favour of the setting.

Then the legs became a cabinet

The second half is the interesting one. A leg is judged alone, but it is not run alone; the legs run together on the accounts Kingdom Portfolios owns and trades for itself, and the group has properties none of the individuals have. We call the assembled group the cabinet, and two things about it surprised us.

The first is that a leg which looks poor by itself can be a good addition. An instrument that measures worse than its neighbours can still earn its seat if it is awake when the others are asleep, because a cabinet that can only act during one session spends most of the week unable to do anything at all. Coverage is a property of the group, and you cannot see it by grading members one at a time.

The second is how the cabinet chooses. The intuitive way to allocate the few open positions a cabinet is allowed at once is to give them to whatever has been performing best. In our own measurement across hundreds of legs, in our own sample, past return did not merely fail to predict the next period; it *anti*-predicted it, while drawdown behaviour was the one thing that tended to persist. That is what we found in that data, not a property of markets we can promise forward. So the cabinet does not rank on recent winnings. It scores on how a leg behaves when it is wrong, and the open-position slots go to what that score says rather than to what the equity curve flatters.

That is what makes the cabinet improve in evidence rather than in size, on our own accounts. Every season it gains another stretch of honest out-of-sample measurement about each leg, legs that only ever looked good in one window fall out of the scoring, instruments are added for coverage rather than for their curve, and the same small number of open positions is spent on better-evidenced candidates. None of that is a claim about what those positions will do. The number of trades does not have to grow; what can grow is how well qualified the candidate standing in each slot is.

Then it went through the gauntlet

This is where most home-built algos stop, with a backtest and a nice curve. We do not count a backtest as evidence. Conduit had to beat a random twin: the identical exits, the identical sizing, the identical costs, and an entry chosen at random instead of by the rule. If the rule is doing anything at all, it has to beat the twin, and it has to do it across more than twenty years of data, not the last good year. The twin test is its own note. The short version is that this single test has killed more of our own ideas than every other check combined. Conduit is one of the few we have not been able to eliminate with it so far, which is a statement about what has not happened yet rather than a claim that it works.

Then it broke, and we said so

A few weeks into the forward work we found a leak: one of Conduit's filters was reading a piece of higher-timeframe information a bar earlier than a live trader could have known it. Every number we had produced for three weeks was contaminated. We fixed it, re-ran everything, and wrote it up in the leak. We mention it here because it is part of how Conduit was built. An algorithm that has never been caught out has not been tested hard enough.

What Conduit is today

A one-trade-at-a-time logic, run as a cabinet of per-instrument legs, still standing after the twin test and corrected when the twin caught the plumbing lying. It is also the seed of Aqueduct, which is Conduit run a particular way. Neither is for sale and neither trades anyone's account but our own. What we share is the way it came to exist, because that way is available to any steward with one honest rule and the patience to test it properly.

Kingdom Portfolios LLC is an independent education publisher. It is not registered with the CFTC or the NFA, is not a Commodity Trading Advisor, does not offer or manage investments, and does not trade anyone else's capital. Nothing here is investment advice, a signal, an offer, or a solicitation.

No performance results of any kind are presented in this note and none should be inferred. Where this note refers to our lab, our harness or our measurement, it means simulated research on historical data. Hypothetical performance results have many inherent limitations: they are prepared with the benefit of hindsight, they do not involve financial risk, no hypothetical record can completely account for the impact of financial risk in actual trading, and no representation is made that any account will or is likely to achieve results similar to anything described. Past performance is not indicative of future results.

The algorithms described trade only Kingdom Portfolios' own accounts. They are not for sale, not licensed, and not run for anyone else. Trading forex and futures carries substantial risk of loss, including loss of the whole account. Education only.