Gradient training
LibTorch runs model training and evaluation without moving the runtime into a separate language stack.
ENZO Trading System Technical reference
One system for models, capital, and execution.
ENZO-TS connects model training, exchange-aware simulation, signal distribution, portfolio allocation, risk controls, and account execution through a shared runtime.
C++20
LibTorch
REST + WebSocket
TCP/IP + IPC
01
Research
Data → train → simulate → evaluate
02
Portfolio
Rank → gate → size → allocate
03
Execution
Signal → order → state → reconcile
TSP
TSM
TSL
TSC
Model work
Gradient training, genetic algorithms, neuroevolution
Risk path
Allocation, gating, entry tolerance, stops, retry
Venue path
Market data, account events, orders, reconciliation
01
System architecture
A process boundary for each responsibility.
Signal generation, network distribution, portfolio orchestration, and account execution run as distinct process types. Shared libraries carry models, exchange access, simulation, reporting, configuration, and networking underneath them.
Input
Market data
Streams + cached candles
TSP
Signal generators
One process per market
TSM
Network hub
Signal distribution
TSL
Orchestrator
Ranking + allocation
TSC
Account clients
Orders + reconciliation
Output
Venue APIs
REST + WebSocket
Shared libraries
Models
Indicators
Simulation
Exchange
Storage
Reporting
Networking
01 Account credentials stay inside the execution clients.
02 A market process can be supervised independently of the portfolio.
03 The network hub can distribute one signal stream to multiple clients.
04 Client and exchange order IDs persist through restart reconciliation.
02
Model development
Models train and evaluate inside the runtime.
LibTorch carries training and model evaluation in C++, with PyTorch alongside it for research. The framework supports conventional gradient training as well as population-based search through genetic algorithms and neuroevolution.
-
01 / Prepare
Historical store
Sub-minute candles, local cache, custom indicators
-
02 / Define
Model families
Strategy framework and candidate network structures
-
03 / Train
Search parameters
Backpropagation, genetic algorithms, neuroevolution
-
04 / Simulate
Exchange behaviour
Orders, fills, commissions, and trading costs
-
05 / Evaluate
Compare candidates
Out-of-sample statistics and leakage controls
Population search
Genetic algorithms and neuroevolution operate over populations of candidate networks.
Evaluation discipline
Out-of-sample evaluation and leakage controls separate development data from validation data.
03
Simulation + evaluation
The simulator models orders, not ideal closing-price fills.
Historical signals pass through exchange-aware order behaviour. Commissions, trading costs, positions, and trades are persisted before the reporting layer calculates return, risk, and monthly statistics.
InputCandles + signals
OrdersVenue rules + costs
StatePositions + trades
OutputEvaluation report
Portfolio runs span multiple markets so allocation behaviour is evaluated inside the run.
04
Portfolio + risk
Risk is expressed as executable controls.
The orchestrator filters and sizes candidates before account clients receive an instruction. Entry, position, and recovery controls continue along the execution path.
-
01
EligibilityMarket gating
Trading switches and active-market limits
-
02
CapitalSize + allocate
Equal, linear, fixed, or manual modes
-
03
EntryDrift tolerance
Skip a fill after price moves beyond tolerance
-
04
PositionStop paths
Stop-loss with a fallback stop path
-
05
RecoveryCooldown + retry
Controlled follow-up after a failed action
Portfolio boundary
Active-market limits are applied before weight is assigned.
Position boundary
Configurable leverage and per-market switches constrain account instructions.
Operational boundary
Persistent state, reconciliation, and process restart cover interrupted runtime paths.
05
Interfaces
Research and controls remain inspectable.
Desktop interfaces configure portfolio simulations and expose market selection, allocation, sizing, drawdown thresholds, and per-market results in one workspace.
Market selection, weight mode, drawdown threshold, and sizing rules on the left; per-market simulation panels on the right.
Open full size ↗