Skip to content
vukem-trading

Reference projects

Jesse: patterns for vukem-trading

Source: jesse-ai/jesse (MIT) and docs.jesse.trade. Researched 2026-11-04. Jesse is a Python crypto framework; items marked unverified were not confirmed in the docs.

Source: jesse-ai/jesse (MIT) and docs.jesse.trade. Researched 2026-11-04. Jesse is a Python crypto framework; items marked unverified were not confirmed in the docs.

Architecture#

Modes: backtest, live/paper, benchmark and optimize. A Research API exposes candles, backtesting and optimization to scripts and notebooks. Storage is PostgreSQL with optional TimescaleDB. Settings live in .env, and a change does not affect running sessions. Internal module boundaries are unverified.

Strategy API#

  • Hooks: should_long, should_short, should_cancel_entry, go_long (sets entry self.buy, self.stop_loss, self.take_profit), go_short, update_position, before, filter_*, watch_list, trading_hours.
  • Context properties: price, balance, available margin, fee rate, leverage, position, average entry price, liquidation price, funding rate, and the mode flags is_backtesting, is_papertrading and is_livetrading, so one strategy runs unchanged in every mode.
  • Hyperparameters are declared as data: name, type (int, float, categorical, bool), min, max, step, options, default; read via self.hp[...].

Backtest realism#

Every candle in the range is simulated, with fees, order types, partial fills, leverage and shorts as first-class. Docs claim multi-timeframe runs are "without look-ahead bias" but do not document the mechanics. Fee and slippage configuration, warm-up and survivorship handling are unverified.

Metrics#

Total closed trades, net profit, max drawdown, max underwater period, annual return, expectancy, average win/loss, win rate (long/short), average holding time, Sharpe, Calmar, Sortino, Omega, streaks, positive/negative months, total fees, and a Buy and Hold benchmark with no leverage and no fees. No JSON schema is documented.

Optimization and overfitting#

  • Optuna with Ray for parallelism. The default objective is Sharpe; Calmar, Sortino and Omega are selectable.
  • Recommended split: training 60-70%, testing 15-20%, validation 15-20%. Validation is used only to choose among finalists. Warning signs: good train with bad test results, and very few trades in the test period.
  • Monte Carlo (trade-order shuffling and candle perturbation) and bootstrap significance testing are offered as guards against overfitting.

Risk controls#

The docs describe liquidate(), available_margin and liquidation_price, but no max-drawdown stop or sizing rules. The main safety statement is a risk disclaimer.

Recommendations for vukem-trading#

  1. Declare strategy parameters as data (name, type, range, step, default) so the optimizer, backtester and UI share one source.
  2. Use a three-period split (train, test, validation); touch validation only for the final choice, and enforce a minimum trade count per run.
  3. Add Monte Carlo and bootstrap significance as report steps.
  4. Put the run mode on the strategy context so identical code runs in backtest, replay and paper.
  5. Always show a Buy and Hold baseline without fees or leverage.
  6. Apply config changes only at the next session start.