What we've built, what it's doing, and where it actually has edge.
Markitel is a real-time multi-asset signal platform with an integrated automation rail, an MT5 broker bridge, and a daily LLM analyst → critic → auto-tuner loop. The infrastructure is unusually thorough for a stage-zero company. The honest open question — the one we want your sharpest read on — is whether classical multi-indicator TA on liquid markets can clear costs at all, and if not, which of the additions on our roadmap actually moves the needle.
Numbers above are real-data snapshots from production Supabase. The v2 rail isn't magic — its advanced features (ranker, regime, bandit, adaptive sizing/exits, reentry) are all flag-off. v2 wins the slice mostly because it didn't take London-session signals during a bad London week. Audit detail in section 5.
What Markitel is, in one screen
A signal-first trading product. Users see ranked, calibrated trade ideas across FX/metals/indices/crypto, can copy them manually, or enroll them in an automation agent with their own rule set (per-rule confidence floors, asset filters, schedule, risk & cooldown caps). Paper trading is the default; live trading routes via an MT5 Expert Advisor bridge. Everything is paper-gated, promotion-gated, and circuit-breaker gated before any user ever risks live capital.
The signal pipeline, end-to-end
A signal flows through six bounded stages before a user can see it; a candidate trade flows through nineteen-plus risk gates before any order fires. Every stage and every gate writes an audit row (signal_gate_events, automation_decisions.gates_evaluated[]). Reversibility is a first-class property — every model change ships behind an env flag, a parameter row, or a model-version pointer.
signals_shadow.signals, factor scores persisted, narrative cached, gate events logged.The honest claim isn't that the model is exceptional — the literature is clear that classical multi-indicator TA on liquid markets has near-zero edge after costs. The honest claim is that the rails are. Each step in 02–06 is reusable for any future model: drop a learned classifier into 02, the shadow rail in 04 measures it, the critic in 06 rejects bad parameter changes, and the live publish in 05 is a single config flag away.
The engine, what it actually computes
Confluence — the 10-factor classifier
A weighted heuristic vote. Min-confluence threshold defaults to 52; per-asset overrides live in engine_parameters.
| Factor | Weight |
|---|---|
| Trend structure (HH/HL or LH/LL) | 20% |
| EMA ribbon alignment (8/21/55/200) | 15% |
| RSI regime | 12% |
| MACD histogram & signal cross | 12% |
| ADX strength | 10% |
| Volume / relative volume | 10% |
| Bollinger band position & squeeze | 8% |
| Stoch-RSI cross | 5% |
| S/R proximity | 4% |
| Liquidity-sweep / wick reject | 4% |
Six setup types: TREND_CONTINUATION TREND_BREAKOUT MEAN_REVERSION MOMENTUM_SURGE DIVERGENCE_REVERSAL SQUEEZE_BREAKOUT. Each can be disabled per-asset by the auto-tuner.
Validation toolchain
- walk-forward
- Replay last 60d against any proposed parameter delta; reject if winners drop faster than losers.
- calibration
- Isotonic regression (PAV) on confidence → realized hit rate, fit monthly on 90d closed signals.
- stats
- Paired bootstrap (p<0.05) + Bonferroni, applied by the critic before any promotion.
- sanity
- Rule guards (n≥8 per asset, 7d anti-thrash, hard envelopes) then a second-LLM skeptic.
- shadow
- Every model change runs into
signals_shadowfor ≥ 14 days before the live pointer flips.
Self-tuning loop (daily)
- 00:10 UTC
engine-analyst— Opus 4.7 reads 14d cohort + factor weights + bounds; returns markdown analysis + JSON recs. - 00:30 UTC
engine-auto-tune— rule gate → challenger critic → walk-forward → bounded delta (max 2 pts/day). - every 6h
engine-critic— second-pass review of analyst recs. - monthly
calibrate-confidence— re-fit isotonic regression.
The automation rails — v1 today, v2 in promotion
Users do not auto-trade by ticking a box. They subscribe to skills (rule rows) that each carry an explicit match_dsl (which signals?) and action_dsl (what risk, what exits?). Every candidate goes through 19+ ordered gates and writes an automation_decisions row whether it's accepted or rejected.
v1 — production rail
- cron
automation-tick · */2 * * * *- partition
- users with
agent_version IS NULL OR 'v1' - dispatch
- first-match candidate → gate stack → execute
- state
- fully live, ~1.7k closed trades / 30 d
30-day v1: −$50,886 · 38.3 % WR · profit factor 0.56. This is the slice we need to fix.
v2 — parallel test rail flag-gated
- cron
automation-tick-v2 · * * * * *- partition
- users with
agent_version = 'v2' - features
- ranker · regime · bandit · adaptive sizing · adaptive exits · reentry · dynamic cooldown
- flag state
- master ON · all 7 sub-features OFF
- guards
- paper-gated 14 d · ≥ 100 closed trades · promotion-criteria · admin live-flip · circuit breaker
v2 (7d): +$5,550 · 52.7 % WR · pf 1.46. Caveat: the lift is mostly cohort & session selection, not yet the smart features. They're still flag-off.
Risk-gate stack (in order)
Defined in lib/risk/automation-gates.ts. First reject short-circuits. All evaluated gates are recorded with latency in automation_decisions.gates_evaluated[].
bridge_orders + trades.Live metrics — what the production data actually says
7-day realized P&L (snapshot 2026-05-11)
| Slice | Closed | WR | Net P&L | Avg / trade | Profit factor |
|---|---|---|---|---|---|
| All rails | 877 | 43.4 % | −$30,067.86 | −$34.28 | 0.59 |
| v1 | 689 | 40.9 % | −$35,617.90 | −$51.70 | 0.41 |
| v2 | 188 | 52.7 % | +$5,550.04 | +$29.52 | 1.46 |
30-day realized P&L
| Slice | Closed | WR | Net P&L | Avg / trade | Profit factor |
|---|---|---|---|---|---|
| All rails | 1,967 | 39.7 % | −$45,336.43 | −$23.05 | 0.64 |
| v1 | 1,779 | 38.3 % | −$50,886.47 | −$28.60 | 0.56 |
| v2 | 188 | 52.7 % | +$5,550.04 | +$29.52 | 1.46 |
The realized headline is harsh and the floating book makes it harsher: 329 active positions across the 30-day slice with stored floating P&L of −$135,349. Any funded-style equity curve must include unrealized P&L — the realized line alone is misleading.
Per-asset edge — last 60 days (raw signal level)
This is the diagnosis we need pressure-tested. The roadmap (Section 7) restricts live emission to XAUUSD only while everything else moves to shadow-only, then rebuilds the classifier on labels rather than intuition.
Cron map — what runs, and how often
29 scheduled jobs, all UTC. The two that matter most: generate-signals (the engine) and automation-tick-v2 (the agent).
| Path | Schedule | Purpose |
|---|---|---|
| generate-signals | */2 * * * * | Scan + emit new signals |
| update-signals | */2 * * * * | Price ticks, TP/SL transitions |
| automation-tick | */2 * * * * | v1 agent dispatch |
| automation-tick-v2 | * * * * * | v2 agent dispatch |
| automation-position-tick | */2 * * * * | Mark-to-market + exits |
| engine-watchdog | */5 * * * * | Engine-silence detector |
| broker-heartbeat | */5 * * * * | MT5 EA liveness check |
| reconcile-trades | */5 * * * * | Bridge vs trades reconciliation |
| Path | Schedule | Purpose |
|---|---|---|
| engine-analyst | 10 0 * * * | Daily Opus 4.7 review |
| engine-auto-tune | 30 0 * * * | Bounded parameter delta |
| engine-critic | 0 */6 * * * | Second-LLM critic |
| calibrate-confidence | 0 4 1 * * | Monthly isotonic re-fit |
| refit-session-policy | 0 22 * * 0 | Sunday session-quality refit |
| refit-payoff-distributions | 0 23 * * 0 | Sunday TP-distribution refit |
| agent-v2-promotion-eval | 0 6 * * * | v2 paper→live gate eval |
| detect-anomalies | */5 * * * * | Drift & tail-event sniffer |
The honest postmortem — three incidents, three lessons
Trade volume 76× normal · 9 % WR
From 5/day on Apr 28 to 380/day on May 4. 974 closed trades, −$52,483 on a $8,150 paper account. Hour 06:00 UTC alone: 12 trades, −$39,346.
Root cause: Loss spiral on broad-match rules with no post-loss throttle that worked in paper. 96 % of trades after a loss within 30 min also lost (823 within 5 min).
Fixed: per-asset daily-loss cap; gate-cap on max-decisions-per-cron-tick; engine-silence watchdog (paradoxically also caused by this episode).
+$17.9k → +$323 after normalization
BNBUSD & SUIUSD were absent from the tracker contract-size spec table → fell through to the FX default of 100,000. Eight BNB trades displayed +$17,759 but the real economic gain was +$0.18.
Real edge in that day: small clusters in GBPAUD, EURCHF, XAGUSD, US30. The crypto outlier was an accounting bug, not a strategy.
Fixed: crypto symbols seeded into pnl.ts (BNBUSD pip 0.1, SUIUSD pip 0.0001, …). Future-symbol guard rejects unknown crypto from automation until spec exists.
8 same-side trades in 14 min · paper-blind dedup
130 XAU trades / 14 d across two test users · 45.3 % WR · −$2,056. Both rules had empty match_dsl so "High Confidence Only" wasn't enforcing confidence at all.
Root cause: (1) empty match_dsl on broad-preset rules; (2) v1+v2 cron lanes double-firing the same signal pre-coexistence patch; (3) duplicate-detection gate queried only bridge_orders, so paper trades never tripped it.
Fixed: dedup gate now unions bridge_orders + trades. v1/v2 partition tightened. Still open: server-side default min_confidence=70 on empty match_dsl, and dollar-risk cap (not lot cap) per trade.
The pattern across all three: the engineering scaffolding is genuinely modern, but a one-line oversight (missing spec / empty DSL / wrong table union) produces an outsized symptom precisely because the surface area is large. The lesson is to invest in invariants — guarantees that hold across the whole surface — not heroic per-bug fixes.
The 12-week engine roadmap
Sequenced for reversibility. Each phase ships behind an env flag or model-version pointer; each phase can be reverted by the critic if shadow loses to live. No flag-day cutover at any point.
LIVE_EMISSION_WHITELIST=XAUUSD; weighted per-asset caps by WR; engine-heartbeat alert if no signals > 4 h in market hours; calibrated WR fed into analyst prompt.
/admin/engine/dashboard; drift alerts on WR / calibration / regime; weekly auto-retrain with shadow → critic → promotion. Mean-time-to-detect: 8 d → 1 h.
Out-of-scope (deliberately)
- Deep-learning sequence models on raw price. Tabular XGBoost is competitive and 100× cheaper to interpret. Skip until classical ML is exhausted.
- Tick-data / order-book features. Requires a paid feed + a real-time pipeline. Maybe phase 7+.
- RL for execution. Poor sample efficiency on non-stationary data; Kelly + heuristic execution covers 80 %.
- Custom matching engine. We're a signal product, not a venue. Don't blur the line.
Funded-trading pivot — conditional go, not yet
We have a private-preview /challenges page (10k / 25k / 50k / 100k / 200k tiers, 8 % target, 5 % daily / 10 % max drawdown, 80–90 % payout split). We are not launching it publicly. Two reasons.
Why not yet — product
- Paper engine is not yet broker-grade — no quote conversion, no swap accrual on paper close, no stop-loss slippage, no spread expansion on news, no margin-call model.
- Active floating book sits at −$135k on the test cohort. A funded engine must score realized and unrealized together.
- Loss-after-loss is too aggressive: 45-loss streak observed in the 7-d slice. Challenge-grade products need hard throttles.
Why not yet — regulatory
- CFTC enforcement is active in this exact category (Traders Global / My Forex Funds complaint focused on "live funds" misrepresentation).
- NFA registration thresholds (CTA / CPO / FCM / IB / RFED / AP) depend on actual business activity — counsel-reviewed before any public launch.
- FTMO's own terms explicitly frame Challenge accounts as simulated — not real-instrument trading.
The v1 model we'd ship (after gates)
Do
- Simulated capital only · no user funds pooled · no custody.
- Rewards = contractually defined, simulated-performance based.
- Markitel does not execute trades on user behalf; the challenge is a discipline-and-skill layer on top of Markitel intelligence.
- Audited methodology + published payout math.
Don't
- No real-funded accounts, no hedged prop book, no pooled trading capital.
- No unverified trader-count / payout / pass-rate claims.
- No marketing that mixes "Markitel executes no trades" with "funded with firm capital".
Open questions — where we want your hand on the wheel
- Predictive edge. Is multi-indicator TA on liquid markets ever an edge over a 14 d / 60 d window with realistic costs? Or is the only honest path features the literature actually validates (COT, options skew, flows, microstructure)?
- Sample size for promotion. We use bootstrap p<0.05 paired against live on net pips + WR + max DD over 14 d. Is that enough for a flag flip, or do you want 30 d / per-regime stratification?
- Calibrated 65 % WR at 1:2 R:R — realistic stretch, fantasy, or wrong target entirely? Should we be optimizing expectancy at lower R:R, or selectivity at higher R:R?
- Gating vs pricing edge. Most of our recent wins came from not trading the bad slices (London, broad-match, missing specs). How much further can gating take us before the engine itself has to improve?
- Funded layer. Is a simulated-capital discipline product on top of a signal engine a defensible business — or does it inherit FTMO's regulatory tail without the brand equity? Where would you draw the line?
- Auto-tune autonomy. The analyst→critic→walk-forward→tuner loop runs daily. Should the bound on a single day's parameter delta stay at 2 points, tighten, or vary by regime?
Reference index — files you can pull up
Signal engine
- lib/signals/confluence-engine.ts
- lib/signals/live-signal-engine.ts
- lib/signals/shadow-variants.ts
- lib/signals/technical-analysis.ts
- lib/services/signals/engine-analyst.ts
- lib/services/signals/engine-tuner.ts
- lib/services/signals/calibration.ts
- lib/services/signals/walk-forward.ts
- lib/services/signals/sanity-check.ts
- lib/services/signals/stats.ts
Automation & risk
- lib/risk/automation-gates.ts
- lib/risk/user-trade-gates.ts
- lib/automation/match-evaluator.ts
- lib/automation/v2/dispatch.ts
- lib/automation/v2/ranker.ts · regime.ts · bandit.ts
- lib/automation/paper-fill.ts
- lib/automation/paper-cost-model.ts
- lib/automation/floating-pnl.ts
- lib/services/orders/execute.ts
- lib/services/tracker/pnl.ts
Scheduling
- vercel.json (29 crons)
- app/api/cron/generate-signals/route.ts
- app/api/cron/automation-tick-v2/route.ts
- app/api/cron/engine-analyst/route.ts
- app/api/cron/engine-auto-tune/route.ts
- app/api/cron/engine-watchdog/route.ts
- app/api/cron/agent-v2-promotion-eval/route.ts
Reading material
- docs/SIGNAL_ENGINE_ROADMAP.md
- docs/markitel-paper-agent-funded-decision-memo-2026-05-11.md
- docs/agent-performance-alextfx-2026-05-12.md
- docs/agent-performance-xauusd-analysis-2026-05-13.md
- docs/agent-risk-audit-2026-05-04.md
- docs/agent-v2-handoff-2026-05-08.md
Prepared by alextfx · 2026-05-14 · For external trading-signals advisor · Scope show-and-tell briefing, no code or decisions in this document.
All numerics are pulled from production Supabase as of 2026-05-11 (decision memo snapshot) and verified against the analysis scripts in scripts/analyze-agent-performance.js and scripts/probe-trades.ts. File references are live paths in the working tree. This brief is intentionally short on opinion and long on substance — we want the advisor's judgment, not our own re-stated.