Case study · paper-trading system
Desk Commander is an autonomous paper-trading system that scans the S&P 500 nightly, surfaces EMA-pullback candidates, and routes them through a deterministic execution pipeline — paper-only by design. It exists to scaffold a discipline: separating strategy hypotheses from intuition, and making the validation argument visible before any capital is risked. The audience is the engineer who recognizes the strategy class and reads the 60-day quiet window as the thesis, not a delay.
Strategy posture
The edge is a familiar pattern in equities: when a stock in an uptrend pulls back to a recognized moving average, it often resumes that trend within a defined window. Desk Commander’s contribution is procedural rather than predictive — it formalizes which stocks to watch, when to look, when to wait, and what disqualifies an alert before it reaches the operator. Paper-only by design; live consideration is gated behind validation, not enthusiasm.
The architecture is a four-stage pipeline. A nightly scanner reads end-of-day data for the universe and identifies symbols matching the structural setup; an enrichment stage joins fundamentals and recent context to each candidate; a notification stage emits a structured alert; a paper-execution stage records the entry — all under a fixed monthly cost ceiling. Decisions are rule-based and deterministic, not a probabilistic model dressed as a signal. The strategy class is trend-pullback: a stock in a confirmed uptrend, pulled back to a structural support, examined for a measurable resumption window. Parameter values are deliberately undisclosed — the methodology is the contribution; the values are calibrated against private validation data and stay there.
Validation is treated as a separate discipline from strategy expression. The system walks forward on out-of-sample windows it never saw during calibration; results are reviewed at the regime level, not the trade level, to avoid optimizing the noise. Before any live-money consideration, the system runs a 60-day quiet phase: paper execution continues, alerts continue, but no design changes are made — the surface area sits and is observed. If the thesis survives that window, live consideration becomes a separate decision; if it doesn’t, the system stays paper, indefinitely, without ceremony.
Universe
507
S&P 500 + pulse symbols
Cadence
Nightly
Post-close scan
Cost ceiling
$20
Anthropic / month
Validation
60d
Quiet window before live
System architecture
Step 01
Nightly OHLCV pull for ≈507 tickers (S&P 500 + a tight pulse list). EMA-pullback filter runs against the close, before the next session opens. The regime gate decides whether tonight is a trading night at all.
Step 02
Candidates that clear the regime gate get a Claude-graded setup-quality score. Cost-capped at $20/month — discipline is part of the strategy. Aggregate output only; no per-trade public exposure.
Step 03
Telegram digest in terminal-grade typography. Ranked list, gate verdict, regime context. Human-readable, machine-sourced. No app to open, no dashboard to babysit.
Step 04
Alpaca paper API. Position sizing rule, exit logic, journaled to Supabase. Paper-only by design — the 60-day quiet validation phase is non-negotiable before any live conversation begins.
Decisions, in five stages
Stage 01 of 05 — Universe scan
507 names · S&P 500 + a tight pulse list
507 candidates remaining
Stage 1 of 5 — Universe scan. 507 candidates remaining.Signal pipeline
The pipeline is rule-based and deterministic. Each stage has a clear admission criterion and a clear rejection path — most candidates are dropped, on purpose. Tap a stage to see what passes and what doesn’t.
Stage 01 of 05 · Universe scan
Nightly OHLCV pull for ≈507 tickers (S&P 500 + pulse list). EMA-pullback filter runs against the close before the next session opens.
Accepted when
Symbol shows a confirmed uptrend pulling back to a recognized moving average. Structural pattern is clean.
Rejected when
Trend isn't clean, pullback is shallow, or volume profile disagrees — symbol drops out silently.
Regime gate
The regime gate is not a parameter open to optimization. It’s a refusal mechanism: the system declines to act when the world doesn’t match what it was designed for. Three states, deterministic transitions.
Observing
The default posture. The market is mixed; breadth is uncommitted; the system records candidates but emits nothing.
Pipeline behavior
Candidates logged · no alerts emitted
Each pass produces a structured alert, shipped to a private Telegram channel. The rendering below is illustrative — the categorical fields (regime, setup-quality, category) are public; the ticker symbol and parameter values are not.
Notification surface — illustrative
@desk-commander-digest
ALERT-2026-05-13-0345Z regime = active setup-quality = good category = trend-pullback emitted = 03:45 UTC ticker = [private — disclosure boundary] parameters = [private — disclosure boundary] ALERT-2026-05-12-0345Z regime = active setup-quality = fair category = trend-pullback emitted = 03:45 UTC ticker = [private — disclosure boundary] parameters = [private — disclosure boundary] ALERT-2026-05-09-0345Z regime = observing setup-quality = marginal category = (suppressed by regime gate) emitted = 03:45 UTC ticker = [private — disclosure boundary] parameters = [private — disclosure boundary] ALERT-2026-05-08-0345Z regime = active setup-quality = good category = trend-pullback emitted = 03:45 UTC ticker = [private — disclosure boundary] parameters = [private — disclosure boundary]
Real alerts include the ticker symbol and parameter values. Both are private to the operator and never surface here.
The Anthropic spend is capped before any feature decision. The cap is part of the strategy: anything the system needs from an LLM has to survive a $20/month ceiling, or it does not get built. Months that approach the cap surface as a design signal, not a billing surprise.
Cost discipline — $20/mo ceiling
Anthropic spend · monthly
Nov
Dec
Jan
Feb
Mar
Apr
May
Dashed line = $20 ceiling. The cap is decided before features.
Scanner build
Strategy backtest
Walk-forward
60-day quiet
Live consideration
60-day quiet phase
60 days total
No design changes during the quiet phase. Validation earns the right to be considered.
LiveDesk Commander · paper-execution
Awaiting first snapshot…
Regime
—
Validation
—
Alerts
—
today · — lifetime
Equity
—
Open positions
—
Next scan
—
Last scan
—
Status
—
Aggregate signals via dc-status · whitelist-enforced · no per-ticker
What "validated" means here
“Validated” here means something narrower than “the backtest looks good.” The system is validated when its rule set survives an out-of-sample walk-forward and then survives a 60-day observation window during which no design changes are permitted. Only then does live consideration become an open question, and it remains a separate decision — validation earns the right to be considered, not the right to be deployed.
Walk-forward windows are chosen to span at least one regime transition each — a sample that ends in a single market environment would teach the calibration only that environment, and validation is the discipline of not learning from the test. Regime gates are an additional filter: the system declines to act in environments outside the ones it was designed for, even when its surface rules would otherwise generate a candidate. Window length and regime gating are decisions made before parameter values are touched; values are then tuned against the held-out windows, and the gates themselves are not parameters open to optimization. The system is built to refuse, gracefully, when the world is wrong for it.
Disclosure boundary
The whitelist below is the code-enforced disclosure boundary. Adding a field to the left column requires a PR + operator review; the right column is never crossed.
Disclosed · whitelist
Not disclosed · private
The boundary is the same discipline as the strategy. Refusing to leak is the work.
End of case study