Not a hundred half-built strategies. One validated edge.
Mercurio is one edge surrounded by serious engineering: a regime-gated trend-following strategy, a veto-only AI filter, risk limits enforced in code, a self-reconciling order lifecycle, and a backtesting lab that refuses to flatter itself. Here is the full catalog.
Eight systems, working as one
Every capability exists to protect or sharpen the single edge. Nothing here is decoration.
One validated edge, run with discipline
A dozen strategies were built and backtested. Only one survived honest testing, and it runs at full allocation.
- Single edge: trend following on 1-hour bars — no hundred-strategy promises
- Triple EMA stacks (12/48, 24/72, 48/144) confirm direction across timeframes
- ADX floor near 22 confirms a trend worth riding before any entry
- Deterministic signals: the same data always produces the same decision
Trade the market it was built for
The engine only acts when the broad market confirms a sustained uptrend. Otherwise it waits in cash.
- S&P 500 must close above its 50- and 200-day moving averages
- Ten consecutive confirming sessions are required before longs open
- Long-only by design — no shorting the strategy cannot reliably trade
- Cash in bear, choppy, and unconfirmed regimes — doing nothing is a position
Intelligent screening, as a veto only
A machine-intelligence layer can reduce conviction or stand aside, but it can never increase size. Technicals remain the edge.
- Sentiment screening on each candidate before an entry is allowed
- News, event, and earnings-proximity awareness as a risk filter
- Veto-only: it can skip a trade or cut conviction, never amplify it
- Graceful degradation — if the AI layer is down, the engine trades on technicals alone
Limits enforced in code, not in hope
Every guardrail lives in the risk manager and runs on every trade. Position-size multipliers only reduce, never amplify.
- Maximum 1.5% of capital at risk per trade — a stop on every position
- Circuit breakers at 5% daily, 7% weekly, and 15% drawdown loss limits
- At 15% drawdown: close all, then a 5-day cooldown before re-entry
- Caps of 25% in any single name and 55% in any single sector
- Loss-streak scaling: three losses halve size; two wins restore it
Every order tracked from request to fill
Entries, exits, and protective legs are managed through their whole life — never fire-and-forget.
- Bracket orders held good-till-cancelled so stops survive overnight
- Stuck-order retry: an unfilled exit is re-attempted every cycle until done
- Partial-fill handling reconciles the position to what actually executed
- Buying-power validation before any order leaves the engine
- Paper-mode lock enforced in code — no live capital during validation
A simulation that refuses to flatter itself
Realism is the point. The lab models the costs and gaps that destroy paper-perfect returns.
- Realistic execution costs: 0.15% slippage per side, roughly 0.3% round-trip
- Gap-through-stop fills — when price gaps past a stop, the fill is at the open
- Walk-forward analysis separates in-sample tuning from out-of-sample proof
- Deterministic and reproducible: same config, same window, same numbers
Built to survive its own restarts
A hard-won lesson made permanent: state that affects trading decisions must outlive a restart.
- Alpaca for market data and paper brokerage execution
- State persistence across restarts — a killed strategy stays killed
- Daily reconciliation of the local ledger against the broker of record
- Health checks and circuit breakers on API failures keep the engine safe
Watch every decision as it happens
A real-time dashboard turns the engine's reasoning into something you can read at a glance.
- Real-time dashboard streaming live engine state over Redis pub/sub
- Live P&L with color-coded, context-rich performance metrics
- A trade journal logging every signal, filter decision, and order
- A regime status banner showing why the engine is trading — or waiting
What this machinery produced in simulation
The same live configuration, replayed over 2024-06-01 to 2026-06-01 on $25,000 of paper capital with realistic execution costs.
These figures come from a historical simulation on 102 symbols with $25,000 of paper capital — they are not live trading results and not a forecast. Over a full five-year cycle that includes the 2022 bear market, the same strategy was net negative (roughly -23%, negative Sharpe, drawdown above 50%). It is a bull-market strategy, still in a 12-month paper-trading validation — not trading live capital.
See the catalog in motion
Walk through the full signal-to-trade pipeline, or read the source that makes every one of these capabilities real.