- Add _to_wine_path(): converts Unix paths under Wine prefix to
Windows-style (e.g. ~/.wine/drive_c/... → C:\...)
- Set subprocess cwd to Editor's directory (editor.parent) instead of
dest.parent — MetaEditor is a Windows EXE and its CWD matters
- Use _to_wine_path() for /compile:"..." argument so paths use correct
drive-letter format under Wine
- Apply both fixes to cmd_compile() and cmd_check()
Introduce a configurable sort priority for Set A and Set B in the
`outliers` subcommand, replacing the single-criterion "Result desc".
New CLI flag:
--sort ABBR_LIST comma-separated metric abbreviations
(default: R,EP,PF,RF,SR,P,DD,C,T)
Abbreviations: R=Result, P=Profit, EP=Expected Payoff, PF=Profit Factor,
RF=Recovery Factor, SR=Sharpe Ratio, C=Custom, DD=Equity DD %, T=Trades.
Equity DD % sorts ascending (lower is better); all others descending.
Unmentioned abbreviations are appended at default order.
Two related improvements to the optimization-report parser:
1. Generic column typing — drop the hardcoded 'InpUseNewsFilter'
branch in parse_passes. Now the SpreadsheetML first body row's
<Data ss:Type='String'> marks a column as string (boolean Inp*
rendered as 'true'/'false' stays str); everything else is numeric
(Int64 when whole-numbered, float64 otherwise). No Inp* name is
referenced, so the parser handles any EA's parameter naming.
2. Generic param detection — _param_cols now returns every column
AFTER 'Trades' by position, not by Inp* prefix. Real exports don't
always use the Inp prefix; per user, 'Trades 以后的列数至少有一列,
但数量不定, 它们都是加入优化的输入参数'.
3. New 'outliers' subcommand — per-pass z-score scan over the 8
performance metrics (Result, Profit, Expected Payoff, Profit
Factor, Recovery Factor, Sharpe Ratio, Custom, Equity DD %).
Splits passes into two disjoint sets sorted by Result desc:
- Set A: at least one metric with |z|>=σ in the favourable
direction (higher-is-better metrics: z>=+σ; Equity DD %
uses z<=-σ because low DD is good).
- Set B: no performance-metric outlier.
Both sets EXCLUDE passes whose Trades count is itself a low-side
outlier (z<=-σ) — too few trades to trust. Excluded list shown
separately. Default σ=2, top 10 outliers / top 5 normal;
--sigma / --top-outliers / --top-normal / --json flags.
Output prints a per-metric mean/std/±σ reference table, then
each record split into Metrics group + Params group with the
outlier σ values annotated. Style mirrors the windows subcommand
in parse_tester_report.py.
Verified on jobs/246753 (432 passes, OneShotGold XAUUSD H4): 44 Set A
passes, 382 Set B, 6 excluded low-Trades passes, all 13 ad-hoc
verification assertions pass.
Split backtest into N equal time slices (left-closed right-open) and
compute the 7 core metrics per window: Profit, EP, PF, RF, Balance DD
Rel%, Trades, Sharpe. Each window gets an outlier flag based on per-
metric z-score (|z|>=2 = notable, |z|>=5 = extreme). N=1 runs a full-
period cross-check vs the HTML report.
Key changes:
- Add compute_windows / compute_window_metrics / print_windows /
windows_comparison functions, CLI subcommand 'windows'
- pair_trades now exports gross_pnl/entry_costs for MT5 GP/GL split
- compute_gross_profit_loss: MT5 accounting (entry costs always to GL)
- _balance_dd_relative: max relative DD (STAT_BALANCE_DDREL_PERCENT)
- _sharpe_ratio: textbook (AHPR-1)/std_HPR formula, 365-day year
- Help text with examples for both --help and windows --help
- verify_sl_tp_formulas.py: localize all output labels to English
- AGENTS.md / SKILL.md: document windows subcommand conventions
Docs: 5 of 7 metrics exact for N=1 (Profit, EP, PF, Trades exact;
RF/BalDD% are approximations due to balance-only reconstruction;
Sharpe uses textbook formula diverging from MT5's 22.92)
New script skills/mql5/scripts/parse_optimizer_report.py reads the
SpreadsheetML export from MT5 Strategy Tester optimization (one row per
parameter pass, plus a <DocumentProperties> environment card).
Companion to parse_tester_report.py. --analyze adds:
- Strategy environment card (EA / Symbol / Period / date range from
Title; deposit / leverage / server / MT5 build from DocumentProperties)
- Orthogonality check (actual passes vs expected cartesian product)
- Parameter effect ranking (effect_ratio = spread / global std) and
dead-parameter detection (per-group mean range < 1% of max)
- Dead boolean parameter detection (bit-for-bit identical true/false
groups across all key metrics)
- Duplicate metric vector counting (>30% usually means a dead param)
- Top-5 best passes by Profit / Profit Factor / Recovery Factor / Custom
- Trade count distribution with daily rate and correlations vs profit
and drawdown (overtrading / undertrading detection)
SKILL.md adds a new 'Optimization Report Analysis' subsection under
the Backtesting chapter covering environment card, orthogonality,
dead parameter / boolean parameter / duplicate detection, multi-
criteria best-pass selection, parameter effect ranking, trade count
diagnostics, and a reporting template.
Validated against jobs/246753/ReportOptimizer-*.xml (OneShotGold,
XAUUSD H4, 432 passes): correctly identifies InpUseNewsFilter as a
dead boolean parameter (true/false identical on Profit/PF/RF/Trades,
189 of 432 metric-vector duplicate groups), reports the trade/profit
overtrading signal (corr -0.56), and ranks best passes.
Adds pandas dependency (used for groupby/aggregate/corr on the pass
table; the script falls back to dict-list output if needed).
- Remove gap_days_to_end and last_trade_close from analyze_report output
- Add idle_time: total backtest duration minus all position holding times
(includes flat time before first trade and after last trade)
- Add format_duration() helper for timedelta -> HH:MM:SS formatting
- Parse both start/end dates from period string for accurate calculation
- Update print_report() to display idle_time in Holding Times section
- main() now always computes analyze data for text report
- Update AGENTS.md: add scripts docs, jobs/resources dirs, verification method
Period format is "H4 (2024.12.25 - 2026.06.22)" — use regex to extract
the date before the closing paren instead of split+strip which fails on
the trailing ")".
Parse end date from Settings.period ("2024.01.01 - 2025.06.22") and use
it as reference for gap_days. Falls back to datetime.now() only if period
cannot be parsed. Rename output key gap_days_to_now → gap_days_to_end.
Resolution order for MQL5_DIR:
1. MQL5_DIR env var (explicit override)
2. Walk up cwd looking for a dir with Experts/ + Indicators/ children
3. MT5_BASE/MQL5 (Linux/Wine default)
When detected from cwd, also infers MT5_BASE as the parent directory.
- MQL5_DIR override via MQL5_DIR env var (Windows 10+ stores MQL5 in
%APPDATA%, not under the install dir)
- compile/check now pass absolute paths to /compile:"path" instead of
computing relative_to(MT5_BASE) which would fail when MQL5_DIR is
separate
ZhijuCen