daniel/Artifex
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
Artifex/docs/trading_program_understanding_20260817.md
Daniel Maddern 65f0d510ec Add offline-first Trading Studio v0.1
2026-08-17 13:48:50 +07:00

6.4 KiB
Raw Blame History

Trading Program Understanding

Scope

This report records read-only archaeology of S:\PycharmProjects\SquadWatch, the Spark replay workspace, and the Pi service state. No exchange action, credential access, service restart, or database mutation occurred.

Current Decision

CONFIRMED: no existing strategy qualifies for Shadow Mode from the latest Spark evidence. The latest 45-day selection/final comparison selected 50 families and recorded survivor_count: 0. This is a successful falsification result, not a reason to loosen the gate.

Data

Status Evidence
CONFIRMED Spark snapshot /home/daniel/hyperscalper-replay/binance_btcusd_2m_455d.csv has 327,599 BTCUSD 2-minute OHLCV rows with timestamp,open,high,low,close,volume. Its observed range is Unix 1743840960 through 1783152720.
CONFIRMED Spark has a chronologically separated 90-day OOS snapshot and 45-day selection/final partitions: binance_btcusd_2m_oos_90d.csv, binance_btcusd_2m_oos_select_45d.csv, and binance_btcusd_2m_oos_final_45d.csv. The selection/final boundary is adjacent, not shuffled.
CONFIRMED The newest discovered Spark HyperScalper artifacts were written 2026-08-08. They are not current market data on 2026-08-17.
CONFIRMED Pi logs were updated 2026-08-17, but hyperscalper-live.service and hyperscalper-risk50-paper.service were inactive during the read-only check.
UNKNOWN Current live exchange positions, fills, and account equity were not queried. Trading Studio V0.1 intentionally does not make exchange calls.
UNKNOWN Funding, open interest, order book, tick, and liquidation datasets were not located in the verified Spark replay input. They must not be implied by OHLCV availability.

HyperScalper Architecture

Status Evidence
CONFIRMED The source system is src/hyperscalper in SquadWatch. It contains indicator/search code, paper replay, feed, bot, executor, risk, and Postgres/SQLite journaling components.
CONFIRMED Live topology documented in docs/hyperscalper_pi_deployments.md is feed -> paper/signal authority -> webhook-only live runner -> Hyperliquid executor. Documented initial live controls were single position, 1x leverage, and max notional 20.
CONFIRMED The Pi documentation states a known accounting limitation: bot-managed exits may record local trigger price/PnL rather than the actual exchange close fill. Historical live PnL therefore needs reconciliation before it is treated as canonical.
INFERRED Existing execution controls are useful source material but cannot be trusted as an Artifex execution adapter until restart recovery, idempotency, account-level risk persistence, and reconciliation are audited.

Features And Leakage

Status Evidence
CONFIRMED fast_engine.py, indicators_depth_nb.py, and indicator_library.py provide a large multi-family feature universe. The implementation uses trend, oscillator, volatility, channel, pivot, time, and candle-derived signals.
CONFIRMED The active paper-like replay computes signal at completed-bar close and enters at next-bar open with intrabar TP/SL handling.
BLOCKED The high-throughput search kernel enters at same-bar close while the paper replay uses next-bar open. Search PnL is not execution-equivalent evidence and cannot qualify a strategy.
BLOCKED Search volatility percentile calibration is derived from the full evaluated window in fast_engine.py; this leaks later observations into early threshold calibration. It must be fit on train only and frozen before validation/holdout.
SUSPICIOUS Existing walk-forward code has no documented purge/embargo around position/cooldown boundaries. Future Artifex folds must embargo at least the maximum holding period plus cooldown.

Existing Strategy Evidence

Status Evidence
CONFIRMED Spark report /home/daniel/hyperscalper-replay/reports/oos_45d_selection_final_comparison.json applied its stated selection rule to 50 candidates and reported zero final survivors.
CONFIRMED The selection candidates exhibited strong prior-window PFs, while the displayed final-window results were negative with PF below one. This is direct evidence of regime/search fragility, not a live execution result.
CONFIRMED walkforward_14d_14d.json states it used a read-only transaction and that selection used only fully completed training trades. It is useful historical evidence, but its thresholds and data partition policy must be revalidated by Artifex before reuse.
UNKNOWN The local 14-row SQLite live journal is too small to establish a live alpha conclusion. The documented main 60-day paper database was not present in the inspected local workspace.

Backtester Integrity Assessment

PARTIAL / NOT QUALIFIED FOR PROMOTION

  • Reuse candidate: src/hyperscalper/paper_replay.py because its timing model is closer to intended paper/live semantics.
  • Reject for qualification: fast_engine.evaluate_combo() and the close-only kernel until same-bar and full-window calibration issues are repaired.
  • Required before a future cohort: immutable source hash, chronological discovery/train/validation/holdout split, train-only thresholds, purge/embargo, fixed cost version, and next-open/intrabar conservative replay.
  • Required before Shadow: a fresh data snapshot after the 2026-08-08 Spark cutoff, a successful independent cohort, and an Artifex shadow runner that records intended order, observed market evolution, and simulated fill without exchange submission.

V0.1 Cohort Policy

The first Artifex cohort is bounded to 20 pre-existing, fingerprint-deduplicated strategy genomes. It must run the paper-like replay with frozen data/splits and cost stress. The latest Spark report already proves that the current 50-candidate selection did not survive its final partition. Artifex must preserve that failure knowledge and must not promote any member of that family from those results.

Shadow Readiness

NOT READY. No shadow command is authorized to start yet. Once a new current market snapshot is registered, has passed quality validation, and an offline cohort produces a holdout-qualified survivor, the future explicit command is:

python manage.py trading_shadow --project crypto-hyperscalper --strategy-version <uuid> --dry-run

That command is intentionally not implemented in V0.1 because it would create a false impression that shadow qualification is currently satisfied.