Skip to content
⚠️ Not financial advice. For educational use only.
WikiHigh Level Design

Trading System High-Level Design (2025 Comprehensive Blueprint)

This document provides a full high-level picture of the trading research platform. It is designed to be read by engineers, quants, and LLMs to understand how all components fit together, how to extend them, and how to use the system effectively.

1. Overview

The project is a modular research and backtesting framework for algorithmic trading.

It supports:

  • Backtesting: bar-driven, phase-aware, realistic fills (tick rounding, slippage, commissions/fees), leverage and maintenance margin enforcement.
  • Live trading: via TradeStation live data (with other broker adapters possible).
  • Strategies (Bots): modular bots that implement entry signals, plug into shared risk sizing, and delegate exits to ExitStrategies.
  • ExitStrategies: traditional ATR-based stops, ML-trained RL exits, or hybrid bandit overlays.
  • Research tooling: offline dataset generation, PPO/LSTM model training, and online LinUCB contextual bandits.
  • User Interfaces:
    • GUI backtester (PyQt + PyQtGraph)
    • Headless backtester (CLI + tqdm)
    • Live TradeStation viewer

Goal: provide a reproducible research environment where strategies and exits can be developed, tested, and compared consistently in simulation, then applied in live data contexts.

2. Architecture at a Glance

+===========================================================================================+
|                                ALGO-TRADING-STACK (2025)                                  |
|                        Modular Backtesting & Research Platform                            |
+===========================================================================================+

 Data & Brokers                        Core Runtime                   Strategies / Exits
 --------------                        ----------------               -------------------
 • Yahoo Finance CSVs       -->        • BacktesterEngine      <---   • Bots (CoinFlip, TrendFollowing)
 • TradeStation (live) <-- Adapter(s)  • TradingEnvironment           • ExitStrategy plug-ins
                                       • Indicator Registry           • Trailing ATR / Fixed Ratio
                                       • Event-driven phases          • PPO/LSTM RL exits
                                                                      • Hybrid LinUCB exits

 Runners & UI
 ------------ 
 • GUI backtest (PyQt)
 • Headless backtest (tqdm)
 • Live trading (TradeStation)*

 Outputs
 -------
 • Equity curves & stats • Orders & trade logs • Charts & indicators

3. Core Components

3.1 API Interface

Defines the execution & data contract (APIInterface):

  • connect / disconnect
  • place_order, cancel_order, modify_stop_loss, modify_take_profit
  • get_positions, get_portfolio, get_total_pnl
  • get_asset_data, get_latest_data, get_asset_list

Implementations:

  • BacktesterEngine (simulation):
    • Phase-aware (OPEN / INTRABAR / CLOSE).
    • Orders, fills, slippage, commissions, margin.
    • Equity curves, campaign stats, per-asset DD and expectancy.
  • TradeStationLiveAPI (live data):
    • Fetches minute bars + quotes.
    • Emits bar_updated to UI.
    • Order methods stubbed (data-only).
  • (Alpaca adapter possible but not included).

3.2 BacktesterEngine (simulation)

  • Phases:
    • OPEN: bot runs on open snapshot (H/L hidden).
    • INTRABAR: stops/TP triggered with gap-aware rules and tie-break policy.
    • CLOSE: equity marked, maintenance margin enforced, indicators updated.
  • Margin enforcement:
    • Intrabar worst-case valuation (LOW for longs, HIGH for shorts).
    • Forced liquidation at bar CLOSE.
  • Accounting:
    • Futures tick_size/tick_value/contract_size respected.
    • Realized/unrealized PnL, commissions, fees deducted.
  • Statistics:
    • Per-asset equity series, drawdowns, expectancy, win-rate.
    • Campaign grouping (flat→nonzero→flat).
  • Signals:
    • bar_updated(df), bar_closed(df), bar_advanced(), backtest_finished().

3.3 Trading Environment

Acts as the hub between adapter, bots, and UI:

  • Indicator pipeline: EMA, RSI, ATR, Bollinger Bands (via registry).
  • Signal wiring:
    • At OPEN → compute snapshot, call bot on_bar(env).
    • At CLOSE → recompute indicators only for last bar.
  • Proxy methods: get_portfolio, get_positions, place_order, cancel_order, get_orders, etc.
  • Safe no-lookahead: indicators carried forward so bots never see future data.

3.4 Strategies (Bots)

All bots inherit from BaseStrategyBot, which provides:

  • Risk-aware position sizing:
    • % equity at risk, slippage buffers, commission/fee accounting.
    • Margin caps: qty ≤ available_equity / initial_margin.
  • Eligibility gates:
    • Session hours & holidays.
    • ATR ≥ threshold.
  • Maintenance flattening:
    • Auto-closes positions before 5pm ET.
  • Lifecycle:
    • If flat → call decide_side, compute SL/TP, size, place order.
    • If in position → trail stop with exit_strategy.update_stop.
  • Online learning hook:
    • After each bar, call exit.ingest_trade_log(env) if supported.

Built-in bots:

  • CoinFlipBot — random entries (baseline).
  • TrendFollowingBot — EMA crossback or ATR breakout entries.

3.5 Exit Strategies

All implement:

  • initial_levels(...) -> (stop_loss, take_profit|None)
  • update_stop(...) -> new_stop|None

Variants:

  • TrailingATRExit — ATR trailing stop, no TP.
  • FixedRatioExit — static ATR stop + fixed R-multiple TP.
  • RLTrailingATRExit (PPO) — ATR multiple chosen by PPO policy.
  • SequenceRLATRExit (LSTM) — TorchScript LSTM rolling classifier.
  • Hybrid exits (PPO/LSTM + LinUCB Bandit) — blend prior with online bandit updates.

3.6 User Interface

  • GUI Backtester:
    • Control panel (start/pause/resume/restart, speed slider, rewind/FF).
    • Equity curve plots.
    • Statistics dialog (trades, win/loss, expectancy, DD, fees).
    • Asset viewers with candlestick chart, indicators, order plotting, SL/TP lines, order entry & position panels.
  • Headless Backtester:
    • CLI with tqdm progress bar.
    • Optionally prints bandit stats after run.
  • Live Viewer:
    • Minimal PyQt chart for TradeStation live data.

3.7 Research & Training

  • Dataset generation (generate_ML_SL_Training_data.py):
    • Scans historical bars, simulates all SL multiples, labels best.
    • Exports training CSVs with features [ATR, RSI, EMA, Close, PositionType].
  • Model training:
    • PPO trainer (train_ppo_stop_selector.py) → .zip policies.
    • LSTM trainer (train_lstm_sl_model.py) → TorchScript .pt + .meta.json.
  • Online learning:
    • LinUCB contextual bandits (bandit_overlay.py).
    • Updated from closed trades in real time.
    • Saves/loads per-symbol bandit state.

4. Data & Metadata Management

  • Sources:
    • TradeStation CSVs: TimeStamp, Open, High, Low, Close, TotalVolume.
    • Yahoo Finance CSVs: mapped to [date, open, high, low, close, volume, timestamp].
  • Master timeline:
    • Union of all asset timestamps.
    • O/H/L left NaN on synthetic bars (gap markers).
    • CLOSE forward-filled, volume zero-filled.
  • Config-driven params per symbol:
    • tick_size, tick_value, contract_size
    • initial_margin, maintenance_margin
    • commission_per_contract, fee_per_trade
    • slippage_ticks or slippage_pct

5. Entry Points

  • GUI backtest: 
    PYTHONPATH=. python3 run_backtest.py
  • Headless backtest: 
    PYTHONPATH=. python3 run_backtest_headless.py
  • Live (TradeStation): 
    PYTHONPATH=. python3 run_live.py
  • Generate training data: 
    PYTHONPATH=. python3 bots/generate_ML_SL_Training_data.py
  • Train PPO exits: 
    PYTHONPATH=. python3 bots/train_ppo_stop_selector.py
  • Train LSTM exits: 
    PYTHONPATH=. python3 bots/train_lstm_sl_model.py

6. Advantages of the Design

  • Realism: phase-aware sim with intrabar liquidation & gap-aware SL/TP.
  • Extensibility: bots, exits, and adapters all plug into stable contracts.
  • Unified workflow: same bot/exit code runs in sim and live modes.
  • Research loop: backtests produce labeled data → PPO/LSTM training → hybrid exits with online learning.
  • Reproducibility: random seeds, deterministic backtests (except during bandit exploration).
  • Modular UI: can run GUI, headless, or live charting.

This document serves as the comprehensive blueprint for the project. It is intended to give both engineers and LLMs the full high-level picture of the system: its architecture, extension points, workflows, and usage.

Get StartedBacktester Engine