How it's built

The whole machine.

The engineering behind the edge, drawn out stage by stage. How an idea is found, proven, matured, and finally run live, all on one shared code path.

01 / Overview

The shape of it.

Find an edge, prove it survives, mature it, run it. Four stages, one pipeline. The sections below zoom into each.

Find an edge, prove it survives, mature it, run it. The sections below zoom into each stage.

02 / Pipeline

How it's built.

A layered onion. Each strategy, broker, and data source is an outer-layer addition that slots in without rewriting the core. Same code path drives backtest replay, paper trading, and live execution.

Backtest path

Live path

That same code path runs the discovery search detailed below.

03 / Execution

How a trade fires.

Filtered names are watched cheaply, on bars, all session. The engine only arms the live tick tape when one approaches the zone it is waiting on, so it listens at full resolution for a fraction of the day, not all of it. A quality-gated trigger fires the entry, and from there the trade either runs to an exit or stops out.

04 / Discovery

How a new edge gets found.

Finding an edge is less a pipeline than a dig. A machine scout tunnels through the data where intuition points it, and most tunnels dead-end on noise, overfit, or a mirage that fails the tests. Once in a while one opens onto a real edge worth keeping. Only what survives the skeptic's gauntlet (next) earns a place in the refinement loop.

A found edge is still only a candidate · next it runs the gauntlet · into validation.

05 / Validation

Built to be broken first.

Every candidate runs a three-gate gauntlet before it earns capital. A backtest number is only worth the attacks it survives.

Hover or tap a corner of the triangle.

Is the edge real?

  • Ablation. Drop each factor one at a time. Passengers get cut; only load-bearing structure stays.
  • Out-of-sample split. The edge has to hold in years it was never tuned on, both halves of history.
  • Null and baseline. Measured against the broad population and its own complement, not just in isolation.
  • No-lookahead audit. Every feature uses only data available at the decision bar. Future-peeking artifacts get hunted down.
  • No single-cell trust. Coherent trends across every bucket, never one lucky standout slice.

Does the money survive reality?

  • Real sequencing. Re-run through the actual state machine, not an idealized first-touch model.
  • Tick-level fill modeling. Each critical moment of each trade gets re-evaluated against the actual second-by-second tape. Spread, slippage, and stop wick-through measured per fill, not assumed.
  • Behavior-aware exits. The exit reads how a move is acting (its momentum and participation on the tape) rather than leaning on a fixed level alone, so the system rides healthy continuation and steps aside on genuine reversals.
  • Stop robustness. Vary the stop across a grid. A result that needs one exact stop width is an artifact, not an edge.
  • Compounding decomposition. Separate per-trade quality from "big total because many bets compounded," so a large number is deflated to what it actually is.
  • Lived drawdown profile. Losing-streak and underwater-spell distributions, fixed as a tolerance before a streak rather than rationalized during one.

Can it actually run?

  • Compute budget. A test that only runs in a notebook is not a strategy. It has to evaluate across the full multi-year universe cheaply enough to re-run on every change.
  • Feed limits. Live selection rides a market-data feed with rate limits and latency. An edge that needs data the feed cannot deliver in time does not exist live.
  • Selector latency. The backtest is handed the candidate. Live has to find it first, in time to act, or the move is already gone.
  • Capacity. Thin names cannot absorb size. An edge that evaporates past a small notional is capacity-bound, not free money.
  • Concurrency. Many setups can fire at once. Capital and correlation cap how many are actually takeable on a day.

Survivors enter the refinement loop below.

06 / Refinement

How an idea earns its way live.

Nothing reaches real capital because it looked good once. Every change starts as an idea at the center and has to earn its way outward, one ring at a time. Any ring can also throw off a new idea, which re-enters at the center, so the loop never really closes.