How MONSTER works, and the plan

Nine core agents cover three universes (US stocks, crypto, and both mixed) at three risk tiers (low, medium, high). Three councils trade the weighted consensus of the best strategies in their universe. One global simulator trades every market including FX, futures and indices, on paper only, as a source of ideas. Four rogues are born with the combined knowledge of every other agent, have the full risk range, and are each forced to see the market differently: the original rogue (anything the account can trade), a daily-bars swing trader, a contrarian that only fades moves, and a ranker that scores stocks against each other rather than on their own.

Every agent's strategy is a genome: roughly 80 settings covering which signals it weighs (momentum, mean reversion, trend, breadth, credit spreads, the dollar, futures positioning, short interest, options flow, insider dealing, news read by an AI, Reddit attention and more), how it sizes, where it stops, how long it holds and which regimes it sits out.

Evolution, not fitting. Each learn runs a population of candidate genomes through generations of mutation and crossover. Candidates are scored walk-forward: trained on one stretch of history, then tested on the stretch that follows, repeated across up to sixteen windows with a gap between them so no test data leaks into training. A candidate's score is its out-of-sample result, never its in-sample one.

Three exams. Stock strategies must also work on a fresh sample of 100 S&P 500 companies they will never trade (the pool exam) and, for slower strategies, on twenty years of daily history (the long exam). The three scores are averaged. A rule that only works on the 50 names it trades, or only in the last two years, does not pass.

Replacing a champion is hard. A challenger must beat the sitting champion by a margin, win most of the test windows, and not do worse on a separate set of validation windows it was never optimised against. Most challengers lose. That is the point.

Ensembles and immigrants. When the champion and its two best distinct rivals agree, trading their consensus often beats any one of them; the desk adopts the ensemble when it does. Agents also borrow each other's champions between learns, and the rogues borrow from everyone.

Live reality check. Once an agent has twenty closed live trades, its real results are compared with its backtest. A champion that is losing live while its backtest said it should win is demoted, whatever the backtest says.

Every hour the desk ticks: fresh prices, signals and news; each agent decides; orders go to Alpaca's paper account with protective stops held at the broker. 21:15 Monday to Thursday, just after the US close, every agent runs a full learn. Overnight until 12:45 a deep session with more history and more windows runs for as long as it has. 13:15 a light learn before the US open. 15:15 to 20:45, while the market is open, only the benched agents learn - the trading ones are left alone. From Friday night to Monday midday the weekend session runs continuous deep passes. In total each agent gets several learns a day and the whole desk learns around the clock.

An agent whose out-of-sample score is below the bench line (-0.25) may not open new positions; it manages what it holds and keeps learning. Above the line it trades, but at a size that scales with its score: a quarter size at the line, full size at +0.75. Fragile champions (those whose results change a lot under small perturbations) are capped at half size. All of this is on the ledger page, live.

On top sit fixed human brakes that no agent can learn its way around: daily and weekly loss limits per agent, a drawdown kill switch per agent and for the whole desk, maximum position, gross exposure and correlation-cluster limits, and a hard lock to the paper account.

Now - winter 2026/27: paper beta. The desk runs on paper, learns every day, and publishes everything here. Friends and early users link their own Alpaca paper accounts and follow it at their chosen risk level. The purpose of this phase is a long, unedited live record - enough closed trades across enough market conditions to know whether the edge is real. Expect agents to come on and off the bench; expect the stocks book to lead and crypto to lag; expect changes to the learning as the record teaches us things.

Spring 2027: review. With several months of live paper results, the record is assessed properly: returns, drawdowns, how closely followers' accounts tracked the desk, which agents earned their place. Only then is anything beyond paper even considered - and if the record says the edge is not there, that will be published too.

What this is not. MONSTER is not investment advice and the operators are not regulated advisers. Nothing here is a recommendation. No real money is traded through this site.