QUBITRADE | Labs

Applied Quantum Allocation.
Executed on Actual Qubits.

Qubitrade is an active R&D framework bridging reinforcement learning and hybrid quantum-classical algorithms. Currently in Week 5 of live-capital execution, we bypass simulated buzzwords to run mathematically dense QAOA arrays on real physical hardware.

State Month 2 Live Pilot (U100)
Primary Solver QAOA (p=1) via ibm_runtime
Live Hardware ibm_torino (133-qubit)

The Theorem

The Combinatorial Wall

Classical Mean-Variance Optimization functions elegantly in continuous, unconstrained environments. It takes milliseconds on a standard CPU. However, real-world institutional capital is constrained by transaction costs, discrete trade lots, and strict cardinality limits (e.g., "select exactly K assets from a universe of N").

Introducing these discrete, binary constraints transforms portfolio allocation into an intractable Mixed-Integer Non-Linear Programming (MINLP) problem. As the asset universe ($N$) scales, the state space explodes exponentially ($2^N$). Classical heuristics routinely stall in sub-optimal local minima. Qubitrade hypothesizes that Quantum Approximate Optimization Algorithms (QAOA) provide a mathematically superior pathway to search these hyper-dimensional energy landscapes natively.

Classical Breakdown Matrix
  • ×
    Cardinality Penalty: Forcing $\sum x_i = K$ requires exhaustive combinatorial traversal.
  • ×
    Local Minima Traps: Simulated Annealing and CPLEX relaxations fail to capture true global minima in non-convex risk surfaces.

Translation Layer

QUBO Formulation

To execute on gate-based QPUs, financial constraints must be translated into the native language of quantum mechanics. We map the allocation problem into a Quadratic Unconstrained Binary Optimization (QUBO) structure, natively equivalent to an Ising Hamiltonian.

$$ \min_{x \in \{0,1\}^N} \left[ \theta \left( x^T \Sigma x \right) - \mu^T x + \lambda \left( \sum_{i=1}^N x_i - K \right)^2 \right] $$

Expected Returns ($\mu$)

The linear term. Predicted short-term returns generated by our QSVR ensemble. Mapped to local fields (Pauli-Z) on individual qubits.

Covariance Risk ($\Sigma$)

The quadratic term. Represents asset correlation, scaled by risk aversion ($\theta$). Mapped directly to ZZ-entanglement couplings between qubits.

Cardinality Penalty ($\lambda$)

A massive algebraic penalty ensuring the solver selects exactly $K$ assets, forcing invalid portfolios into high-energy (sub-optimal) states.

System Architecture

The Execution Pipeline

PHASE 1 : INGESTION & KERNELS

QSVR Covariance

Classical data streams are engineered into high-dimensional feature maps. We actively benchmark Quantum Support Vector Regressors (QSVR) against classical counterparts to predict covariance matrices based on 88,000+ ticks of rolling data.

PHASE 2 : ORCHESTRATION

RL Control Agent

A reinforcement learning (PPO) agent dictates state action. The RL policy triggers optimization only when mathematical drift exceeds thresholds computed in the current volatility regime.

PHASE 3 : OPTIMIZATION

QUBO / QAOA Solvers

Target allocations are isolated, constructed into QUBOs, and handed to async Celery workers. The workers submit the parametric circuits via IBM Qiskit Runtime, proving viability on real quantum hardware.

Active State

Current R&D Operations

LIVE PILOT

U100 Real-Capital Execution

Currently executing Month 2 (Week 5) of autonomous live operations. The architecture is managing real capital across a 100-symbol universe, grounding theoretical math in actual market microstructure and slippage.

SCALING

U500 Offline Overhaul

Simultaneously scaling offline data pipelines to target 500-asset indices. We are running rigorous constraint sensitivity analyses and testing the agent's stability against massive topological state spaces.

INFRASTRUCTURE

Batched Quantum Jobs

Preparing to introduce new users by aggregating disparate QUBO formulations into batched quantum sessions. This optimizes NISQ-era hardware queue latency and minimizes QPU idle-time overheads.

Model Validation

Institutional Friction & The Counter-Intuitive Alpha

In quantitative finance, an unconstrained backtest is largely fiction. To validate our new reinforcement agent ahead of a Q2 U500 deployment, we intentionally choked the model with severe institutional friction to observe where the logic breaks.

We locked in a strict 30 bps transaction drag across a 252-day out-of-sample holdout (500-symbol universe), running a constraint sensitivity analysis by progressively strangling the daily turnover allowance from 15% down to exactly 5%.

The telemetry revealed a counter-intuitive structural advantage. When heavily constrained, the agent did not degrade. It mathematically adapted, abandoning marginal high-frequency noise to hold higher-conviction, longer-duration states—fundamentally improving its risk-adjusted survival.

U500 OOS Holdout (30 bps drag) 252-Day Window
Turnover Cap Sharpe Max DD Return Calmar
15.0% Max 1.44 -12.80% +32.67% 2.56
10.0% Max 1.59 -11.22% +34.38% 3.08
07.0% Max 1.73 -10.49% +35.99% 3.45
05.0% Hard Cap 2.19 -09.13% +44.41% 4.89
* Benchmark Equal-Weight (n=500): +23.36%. A 2.19 OOS Sharpe over a single bull regime is a statistical anomaly. Absolute metrics will degrade in live multi-regime execution. The institutional value is the relative delta: compressing turnover expanded Sharpe by 52%.
stdout // U500_AGENT_SIMULATION
svc_quant@ml-node-04:/opt/core$ DAILY_TURNOVER_CAP=0.05 TRANSACTION_COST_BPS=30 python manage.py backtest_agent --model champion --holdout-days 252
--- Starting CHAMPION Agent Backtest ---
Loading historical data from /opt/core/data/historical_data.csv...
Backtest period: 2025-02-27 14:30:00+00:00 to 2026-02-27 20:45:00+00:00
[PID 47119] Initializing environment for 500 symbols...
[PID 47119] Successfully loaded pre-computed QSVR predictions.
Model and stats loaded successfully. Starting simulation...
Simulation complete.
--- Backtest Performance Report ---
Initial Portfolio Value: $100,000.00
Final Portfolio Value: $144,410.94
Total Return: 44.41%
Annualized Return: 44.62%
Sharpe Ratio: 2.19
Max Drawdown: -9.13%
Calmar Ratio: 4.89
Avg Daily Gross Turnover: 0.0498
Trade Steps: 251

Authentic Telemetry

Raw Production Logs

These are strictly un-indented stdout traces from our offline training environments and physical QPU executions on March 4, 2026. For OPSEC, critical hyperparameters, active feature vectors, and network paths have been strictly redacted [REDACTED] to highlight only the raw computational scale: massive data ingestion, out-of-sample cross-validation, and grueling topological hardware transpilation.

stdout // QSVR_KERNEL_TRAINING OPSEC Sanitized
svc_quant@ml-node-04:/opt/core$ python manage.py train_model --min-history-days 60 --n-splits 8 --final-holdout-days 60
--- Starting Quantum ML Model Training ---
Building QSVR supervised frame...
[holdout] Reserved final holdout window: days=60 cutoff_utc=2025-12-01 train_rows=253923 holdout_rows=29551
[INFO] 2026-02-28 23:46:19 :: QSVR CV staged search enabled grid_size=[REDACTED] stage1_folds=2 stage1_jobs=2
...
[INFO] 2026-03-01 07:58:47 :: QSVR CV staged search stage2 selected=12/36
[INFO] 2026-03-01 07:58:47 :: QSVR CV start n_samples=253923 n_features=[REDACTED] n_dates=513 n_splits=8
...
[INFO] 2026-03-01 23:32:31 :: QSVR CV progress config=2/2 fold=8/8 elapsed_s=85572.3
[INFO] 2026-03-01 23:32:31 :: QSVR quantum subset eval complete n_samples=2048 n_dates=30 baseline_ic=[REDACTED]
[INFO] 2026-03-02 00:50:33 :: QSVR OOS artifacts written candidate=nystroem_rbf days=456 mean_ic=[REDACTED]
[QSVR] Best params: [IP_REDACTED_GRID_PARAMS]
[QSVR] Saved best model to /opt/core/saved_models/qsvr_model.joblib
[holdout] mean_daily_ic=0.0142 t_stat=2.18 days=60
--- QSVR Model Training Complete ---
svc_quant@ml-node-04:/opt/core$ python manage.py precompute_qsvr --prediction-batch-size 4096
--- Starting QSVR Pre-computation ---
[QSVR] historical_data.csv latest_ts=2026-02-27T20:45:00+00:00 age_s=212639 max_age_s=345600
[QSVR] Predicting in batches batch_size=4096 n_samples=283474 n_features=[REDACTED] dtype=float32
[QSVR] Scaled predictions to nominal dispersion target=[REDACTED] raw_median=[REDACTED]
[QSVR] Saved wide-format predictions to /opt/core/data/qsvr_predictions.csv shape=(573, 500)
--- QSVR Pre-computation Complete ---
svc_quant@ml-node-04:/opt/core$ python scripts/check_qsvr_health.py --target holdout
QSVR Holdout IC health ok: True
Metrics: {'mean_holdout_ic': 0.0142, 'frac_positive': 0.551, 'status': 'PASS'}
stdout // QPU_HARDWARE_EXECUTION (IBM_TORINO)
[INFO] 14:48:20 apps.broker.tasks :: strategic.qaoa.start n_assets=100 risk_aversion=[REDACTED] n_select=5
[INFO] 14:48:23 apps.quantum.backends :: quantum.ibm.connect.ok job=strategic_rebalance backend=ibm_torino
[INFO] 14:48:24 backend_converter :: Control flow if_else is found but dynamic circuits disabled. Excluded from backend Target.
[INFO] 14:49:06 apps.quantum_connector :: quantum.job.start provider=ibm_runtime backend=ibm_torino timeout=2000s shots=2048 reps=1
[INFO] 14:49:10 apps.quantum_connector :: quantum.transpile.strip_rzz job=strategic_rebalance
# Transpilation Passes (Mapping fully-connected QUBO to Heavy-Hex Topology)
[14:50:45: INFO/ForkPoolWorker-2] Pass: HighLevelSynthesis - 95565.55009 (ms)
[14:50:48: INFO/ForkPoolWorker-2] Pass: SabreLayout - 2846.34399 (ms)
[14:51:03: INFO/ForkPoolWorker-2] Pass: CommutativeCancellation - 5853.81770 (ms)
[14:51:12: INFO/ForkPoolWorker-2] Pass: Optimize1qGatesDecomposition - 2086.46750 (ms)
[14:51:25: INFO/ForkPoolWorker-2] Pass: CommutationAnalysis - 2032.99832 (ms)
[14:51:27: INFO/ForkPoolWorker-2] Total Transpile Time - 137287.79340 (ms)
base_primitive._run:INFO: Submitting job using options {'options': {}, 'version': 2, 'support_qiskit': True}
[INFO] 14:51:30 apps.quantum_connector :: quantum.ibm.job.submitted job_id=d6k4dsg60irc73959m30
[14:52:03: INFO/ForkPoolWorker-2] Optimization complete in 173.858941 seconds.
[14:52:03: INFO/ForkPoolWorker-2] Found opt_params [IP_REDACTED_ARRAY].
[INFO] 14:52:45 apps.quantum_connector :: quantum.job.end job=strategic_rebalance success=True total_s=219.778522
[INFO] 14:52:45 apps.broker.tasks :: strategic.qaoa.success backend=ibm_torino n_selected=27
[INFO] 14:52:45 celery.app.trace :: Task _execute_strategic_update succeeded in 265.277s

Research & Development Collaboration

Qubitrade is strictly an experimental PoC. We are not offering commercial SaaS or managing capital. We welcome dialogue with quantitative researchers, deep-tech infrastructure funds, and quantum hardware foundries interested in algorithm validation.