Sigil's Bullpen: Paper Trading Goes Live

Sigil can rip through hundreds of strategies against years of historical bar data in a few minutes. That’s great for the lab, but historical performance is only ever a hypothesis. The market has to render the verdict.

So today, May 1, we kicked off the next phase: a paper trading simulation, running live against real market data, against real timing, against real slippage assumptions. No real capital yet. Just the strategies, the data feed, and a clock that doesn’t care how clever your backtest looked.

What the Bullpen is

We’re calling it the Bullpen. It’s where new strategies get sent before they’re allowed near anything that matters.

When a strategy survives Sigil’s offline sweep — clears whatever performance, drawdown, and stability gates we’ve defined — it doesn’t graduate straight to production. It gets sent to the Bullpen. There it runs in paper-trading mode against live data, on the same cadence and with the same execution assumptions it would have if it were live. We collect the trades, the signals, the misses, and the moments where the strategy did something the backtest never asked it to do.

The point isn’t to find a winner. The point is to find what breaks. Live data is noisier than historical data. Latency is real. Bars print late, get revised, or arrive with surprises that a clean parquet file never had. A strategy that looks pristine in backtest can fall apart the first time the feed jitters. We’d rather learn that in the Bullpen than learn it with capital on the line.

The May 15 cut

We’re letting this initial run go until May 15. Two weeks is short enough to get a fast read, long enough to catch a few different market regimes, and short enough that a bad strategy can’t quietly rack up a phantom track record we’d later be tempted to trust.

When we cut the runs, the analysis starts:

• Which strategies’ live behavior matched their backtest, and which drifted?
• Where did the Bullpen subroutines themselves get in the way — bad fills, missed bars, late signals, anything that’s our engineering problem rather than the strategy’s fault?
• What’s the right next gate for strategies that pass the Bullpen, and what’s the right way to retire ones that don’t?

The Bullpen subroutines will get the same treatment Sigil’s processing engine did: profile, find the hot paths, fix them, and tighten the loop until the difference between paper and live is mostly market noise rather than infrastructure noise.

Where Sigil is headed

The longer-term goal is for Sigil to be more autonomous than it is today. Continual backtesting on incoming data. Strategies generated, evaluated, and promoted into the Bullpen without a human having to babysit every step. The Bullpen itself watched by the same kind of profiling that made the strategy engine fast, so we can run more candidates through it in less time.

Sigil’s job has always been the same: find strategies worth caring about, faster than someone with a spreadsheet ever could. The Bullpen is the next piece of that. May 15 will tell us how much of it works in the wild, and what needs to be rebuilt for round two.

More to come after the cut.