Private & confidential
Quanthack 2026 · Systematic trading
HELM
An AI-built trading desk, live and armed in the competition.
HELM · Quanthack 202601
The product · live right nowExhibit 1
The desk is live and armed on the competition account, trading and governing its own risk in real time.
The live armed MT5 account, read every 60s from the running loop. Real fills on the simulated venue, at the deployed leverage; hard caps enforced outside the model.
Running now
30-sleeve book, armed
Real fills on the competition account, supervised by a watchdog that relaunches the loop on death and restores risk state after a restart.
Governed
Caps it cannot raise
Halve, flatten and kill latches plus a per-name cap live in deterministic code outside the model. It can de-risk, never loosen a limit.
Provable
Hash-chained tape
Every signal, order and fill is hash-chained and tamper-evident, replayable end to end.
Live demoThe dashboard reads data/run_live.jsonl each poll: account value, return, drawdown, realized leverage and per-name exposure, straight from the armed account.
Source: scripts/run_live.py (armed loop), watchdog.py, dashboard_server.py; engine/tape.py (hash chain)Private & confidential02
Result · the bookExhibit 2
The book backtests at a 3.4 Sharpe, 4.2 held-out, and a 1.3% maximum drawdown, net of costs.
Equal-frequency book, 1x, net of costs, over the window where all validated sleeves are live (Dec 2024 onward).
4.2
Held-out Sharpe across validation folds (in-sample 3.4)
1.3%
Maximum drawdown, net of cost
2.6%
Annualised volatility at 1x
+8.8%
Return at 1x, net of cost; scales with deployed leverage
ExampleThe book blends a gold trend sleeve, FX mean-reversion across EURGBP, USDCAD, AUDUSD, and a crypto flow sleeve, equal-weighted by frequency so no single sleeve dominates.
Source: scripts/portfolio_v3.py, eq_freq book, post-cost at 1x leveragePrivate & confidential03
The filterExhibit 3
Eight tests cut 98 candidates down to a 21-sleeve validated core, since extended to 30 live.
Overfit probability, regime, annual, stability, ablation, decay, trade cost
Drop near-duplicates so the book is not one bet in many guises
- CPCV and PBO to test whether the edge is real or a lucky path.
- Regime and per-year checks: does it hold when conditions change?
- Stability and ablation: do neighbours agree, does each piece earn its place?
- Decay and cost: survive the entry lag and clear cost at its own frequency.
ExampleA London PM-fix gold fade was cut when 4 of 10 cross-validation folds lost money (PBO 0.40 against the 0.30 limit): a textbook crowded-anomaly decay.
Source: scripts/validate_strats.py, verify_survivors.py; data/tournament/final_universe.jsonPrivate & confidential04
Compounding the edge · research in-flightExhibit 4
Mid-competition, the research engine added 10 vetted sleeves in Round 1: a more diversified, more robust book on every performance metric.
| Book metric · before vs after | Before 20 | After 30 | Shift |
|---|---|---|---|
| Sharpe, 1x net | 0.64 | 3.48 | 5.4× |
| Maximum drawdown | −17.4% | −4.8% | ⅓ depth |
| Annualised volatility | 12.8% | 10.6% | lower |
| Net return, 1x | +8.2% | +37.1% | +29pp |
| Top-25 round probability | 4.5% | 7.2% | +60% |
| Forced-liquidation prob. | 2.55% | 0.00% | → 0 |
| Sleeves · instruments | 20 · 9 | 30 · 11 | +10 · +2 |
Live-realization frame: 15-min bars, post-cost, full history, equal-frequency weighted, at the book's 25x optimal gross. A stricter lens than the validated-core figures on slide 3 (daily, held-out CPCV), so absolute levels differ by design; the before/after is one method measured on one shared panel.
- The same /research-strat pipeline that built the core graded a mid-frequency literature wave through the eight-test battery and a shuffled-null gate; 3 are genuine FX mean-reversion diversifiers, and 2 add new instruments (AUDUSD, USDJPY).
- Promotion was gated, not automatic: an add-sleeve Monte Carlo had to show the full set lifts rank odds with no added stop-out before a human re-lock.
The diversification signature
Volatility and drawdown fall while return rises: Calmar 0.47 to 7.7. Breadth across instruments and timeframes, not one new bet, is what lifted the book and removed the tail.
ExampleA vol-seasonal gold-momentum sleeve (XAUUSD 4h) lifted top-25 odds to 5.35% on its own; the AUDUSD mean-reversion pair added breadth in a name the book had never traded.
Source: scripts/scenario_addsleeve_mc.py, scripts/_addsleeve_full_metrics.py, scripts/promote_candidates.py; .claude/skills/research-strat; data/tournament/scenario_addsleeve_mc.json (commit 9c23f0a)Private & confidential05
How the scoring is wonExhibit 5
Hold a steady book, then move risk up or down to climb the rank without getting wiped out.
Over-gamble
All-in at high leverage. A single 5% move against you triggers forced liquidation and elimination.
Where we trade
Enough risk to climb the return rank, never enough to be liquidated. The book sits here.
Too quiet
Run flat to stay safe and the return rank leaves you mid-pack, cut at the next round.
The play
Hold a low-drawdown book, then size up or down against the live rank and the time left to the next cut. Most of the field sits in one of the two losing zones.
ExampleSized so a typical adverse 24h move costs a few percent of equity and never approaches the 30% margin stop-out, the book climbs return rank without ever risking elimination.
Source: engine/tournament_policy.py (variance dial)Private & confidential06
Governing its own risk · the engineExhibit 6
Four layered controls govern the book's risk. It can de-risk itself; it can never loosen a hard limit.
01 · Directional throttle
Caps the one-way bet
When longs and shorts crowd one direction, it scales the whole book down to hold net exposure at the cap, then restores to full as the book re-balances. The systematic answer to a crowded-long day.
02 · Layered loss latches
Halve, flatten, kill
Three escalating, deterministic stops: −3% halves new risk, −6% flattens the day, −10% trails the high-water and kills. Each is a live dial, re-tuned per round without a code change.
03 · Tunable leverage, hard floor
Chase behind, lock ahead
The operator dials gross up to chase the cut when behind, down to protect when ahead, through one switch. The floor is automatic: a trailing stop that ratchets UP with every new high, so a swing never hands the gains back.
04 · One-command cutover
Switch posture, safely
A single human switch re-deploys the live book at a new leverage with a confirmed-single-writer, failure-atomic restart. It never strands an open book without its risk loop.
Live proof · overnight 22–23 June
The engine carried the book from rank 122 into the top 10. The throttle held net at the cap while gross ran a full 25x, and not one latch had to fire.
ExampleBuilt the morning a crowded-long book swung −9% intraday at high gross. The net-exposure throttle now caps exactly that; by the next morning the book ran a full 25x gross at only 4x net, fully diversified.
Source: engine/risk.py (throttle + latches), scripts/set_leverage.py (cutover); live path from data/run_live.jsonlPrivate & confidential07
Live result · Round 2Exhibit 7
Live, behind at rank 122, the book chased safely and climbed into the top 10.
Real account equity from the armed loop, marked every 60s, Round 2 (22–23 June). Net of all costs; the dashed line is the $1M start.
Top 10
Live rank in Round 2, up from 122
+9.5%
Peak equity vs the $1M start
+6.8%
Latest, net of all costs
0
Loss latches tripped overnight
What this provesThe same governed engine that survived a −9% crowded-long day then chased a 100-place gap and held it, running 25x gross at about 4x net, with a trailing floor ratcheting up under the gains. The risk core is what made the chase safe.
Source: data/run_live.jsonl (armed account, 60s marks); rank from the competition leaderboardPrivate & confidential08
When the desk was wrongExhibit 8
We called it wrong in week one, then changed the answer when the cost rules changed.
Week 1
Taker verdict: no edge
At assumed spreads, almost every price signal died on cost. The harness flagged our best-looking ML cell "real: false", a rebate artifact, and we did not deploy it.
16 June
The rules corrected us
The host confirmed real venue cost is a fraction of our assumption and the engine fills makers. Every prior cost-killed verdict was void and re-scored at the measured spread.
Re-derived
Kept only what survives
Re-validated under honest, measured costs through the same eight-test filter. What survives is what we present; what does not, we benched on the record.
Why this matters
What is worth seeing is the system itself: it attacks its own results, backs down when the evidence changes, and logs every reversal on the record. That is the part we built.
ExampleThe one ML cell that cleared every gate (net Sharpe 3.56) was flagged "real": false by our own harness: a maker-rebate artifact (cost share −1.93) that becomes −15.8 Sharpe as a taker. We benched it.
Source: research/tournament/HONEST_FINDING.md; docs/PRD.md (cost model, 16 Jun); research/adversarial_verify.pyPrivate & confidential09
Private & confidential
Recommendation
Run a book sized to clear every cut, and enter the AI desk that built it for Best Technology.
Kit (Keith) So · Principal Researcher & Lead EngineerHELM · Quanthack 202610