Myquant SkyPark: DOT First-Minute Range Breakout with Literal Direction Flags

Derived by the Research Lead from "SkyPark" by Myquant (myquant) https://github.com/myquant/strategy/blob/17595e6bf4a118e1fa87c90bfb0fd78afa69a60b/SkyPark/python/SkyPark.py: Preserved Myquant's literal signal and exit rules; completed unavailable INI settings and adapted instrument, daily initialization, sizing and execution. Distilled translation of SkyPark by Myquant, Apache-2.0, from the supplied commit. Preserve the executable comparisons, first-minute high/low boundaries, flat-only entry and timed liquidation. The Python file specifies OPEN_VOL=5 but reads the instrument, clock times and gap thresholds from SkyPark.ini without defaults. That companion file could not be retrieved; the values below are explicitly platform-selected completions, not claimed source defaults. Use DOTUSDT.BINANCE one-minute bars and trade ticks, UTC daily research sessions, a first-minute interval of 00:00–00:01 and an exit threshold of 23:55. Replace five source contracts with capital-relative exposure. Reinitialize once per UTC day, corresponding to restarting the source for each trading day; the original code itself has no daily reset. Decisions remain bar-triggered, using the latest observed trade price. Market orders execute at the decision bar's close. No price stop or profit target is added. The requested outbox file could not be written because this workspace is read-only.

negative_expectancy: over 2,036 closed trades from 2020-08 to 2026-09 (full history now evaluated after the tick-fallback fix), the Research Lead's DOT operationalization of Myquant's SkyPark loses money after costs. avg_trade_return_pct is -0.69% against a 0.15% BINANCE futures floor. Profit factor is 0.66, Sharpe -1.83 (95% CI -2.63 to -1.08), total return -97.3% and max drawdown 97.5% against a pre-registered 15% limit. Every calendar year is negative. Fees are only 6.7% of gross, so the gross edge itself is negative (signal IC -0.019). DOT buy-and-hold fell heavily over the window, yet this short-only book still lost. Structural note: on a 24/7 market the source's overnight-gap flags carry no information, because pre_close is the previous minute's close. The long flag never fired (0 long trades), and the literal short branch fires almost daily. Optimization has nothing to work with: tunable is empty and all values are frozen. This verdict rejects only the DOT/UTC-session adaptation, not Myquant's original strategy on session-based futures with real overnight gaps. That variant cannot be tested here without such a venue.

Iteration 4 reproduced iteration 2 because the code fix requested in iteration 2 was NOT applied. Apply it now. This change alters trading; a third identical run will abandon the run. DEFECT: should_enter() returns None whenever tick_ready is False. DOTUSDT trade ticks exist only from 2026-04-29 (input_audit first_ns 1777420801092000000). entry_diag shows 190,080 tick-covered bars against 2,989,019 missing. Entries were therefore blocked from 2020-08 to 2026-04, so all 22 trades fall in about 130 days of 2026. That covers about 6% of the 2,207-day history. The rule was never evaluated on the full history. FIX (code only; no fixed contract value changes): 1. In calculate_signal, set self.decision_price = self.last_trade_price if self.tick_ready else bar.close.as_double(). On a 1-minute LAST bar, the bar close IS the last trade at or before the decision time, so this is the same 'latest observed trade price', not a substitution. 2. In should_enter, remove the `not self.tick_ready` condition. Keep `not self.levels_ready`. 3. position_size already uses decision_price, so it will then work without ticks. 4. Keep subscribe_trade_ticks (fixed). Ticks still take priority when present. Expected result: entry_signaled_2020..2025 should be non-zero, with roughly one entry per day on most days. DO NOT CHANGE (literal source semantics, keep as-is): the open_long/open_short flag comparisons, first-minute upr/dwn, flat-only entry and the 23:55 timed exit. NOTE FOR LATER REVIEW: pre_close is the 23:59 one-minute bar close, so first_open vs pre_close almost never gaps by 1%. The short branch's literal '>' therefore sets open_short_flag on nearly every day, and long_trades = 0. That is the literal source behaviour under the platform's UTC/1-minute completion and will be judged on the full-history result. Do not alter it in this iteration. ## Library refinements (from the knowledge library; test them, do not assume them) The library holds the SkyPark source itself, which confirms that entries are decided in on_bar against the latest price. So the analyst's fix (use the bar close when no tick exists) is faithful to the source and must come first. After a full-history run, the library's opening-range material points to three changes: a stop at the opposite edge of the opening range, a wider opening window than one minute, and a range-based trigger in place of the gap flags, which never fire correctly on 24/7 crypto. It also warns that opening-range breakouts often fail to beat costs. 1. [entry] Evaluate entries on full history (drop the tick gate): In calculate_signal set decision_price = last_trade_price if tick_ready else bar.close.as_double(). In should_enter, remove the `not self.tick_ready` condition and keep `not self.levels_ready`. Keep subscribe_trade_ticks so ticks still take priority when they exist. No parameter changes. — In the source, on_tick only stores tick.last_price into self.close. Every decision ('if self.open_short_flag and self.close < self.dwn and 0 == self.hoding') is made in on_bar against the latest price. On a 1-minute LAST bar, the bar close is that latest price. The tick gate is a translation artefact: it limited all 22 trades to about 130 days of 2026, roughly 6% of the history, so the current metrics (Sharpe -0.80, avg trade -1.43%) say nothing about the rule. (source: SkyPark (Myquant) p.1) 2. [stop] Stop at the opposite edge of the opening range: After the full-history run, add a stop in should_exit. Short: exit on a bar close > upr. Long: exit on a bar close < dwn. Then move the stop to break-even once the open profit reaches 0.5%. Keep the 23:55 timed exit and the one-entry-per-day rule. — The QuantInsti opening-range project sets 'First SL ... at the low/high of the first candlestick', then a 'Trade Insurance' move to break-even after a favourable move. SkyPark holds a losing breakout for up to 24 hours with no stop. In the partial run the profit factor was 0.28 and the win rate 36%, so losers are far larger than winners, which is the failure this stop targets. (source: Intraday Low-Frequency Trading Strategy (QuantInsti EPAT project) p.1) 3. [parameter] Widen the opening range from 1 minute to 30 minutes: Set first_kline_time=00:00:00 and first_kline_time1=00:30:00. upr/dwn become the high and low of all 1-minute bars in 00:00–00:30 UTC, and no entries are allowed before 00:30. — The FMZ primer describes 开盘突破策略 as built on 开盘半小时 (the first half hour), which 'often decides the day's direction'. The QuantInsti ORB uses a 5-minute first candle. A one-minute DOT range is a few basis points wide, so almost any noise triggers a breakout. The trade then has to cover the ~0.10% round-trip fee from a near-random entry. (source: 发明者量化交易入门--从基础到实战 p.1; Intraday Low-Frequency Trading Strategy (QuantInsti EPAT project) p.1) 4. [entry] Replace the gap flags with a Dual Thrust range trigger: In a later iteration (once the literal version has been judged), keep the 00:00 UTC open as the anchor. Range = max(HH-LC, HC-LL) over the prior N=2 UTC days. Go long when the price is above open + 0.5*Range, short when it is below open - 0.5*Range, one entry per day, flat at 23:55. — On a 24/7 market, pre_close (the 23:59 close) is almost the 00:00 open, so a 1% gap almost never happens. The literal '>' in the short branch then flags a short nearly every day, and long_trades is 0. The Dual Thrust article describes the same opening-range-breakout idea (开盘区间突破) with a volatility trigger, K1=K2=0.5 in its demo. It notes the system does better in trends and gives false signals in choppy markets. (source: 用Python实现一个Dual Thrust数字货币量化交易策略 p.1) 5. [cost] Require a minimum range width so the move can cover fees: Only arm entries on days when the opening range is wide enough: (upr - dwn) / dwn >= 0.30%, which is 3x the 0.10% round-trip cost. Judge the result by the per-year breakdown in entry_diag, and require a positive net return in at least 4 of the 6 years. — A systematic falsification study of intraday OHLCV momentum on MNQ futures found that most signal families' gross return was below the friction floor. The Opening Range Breakout long cleared costs but failed significance and year-stability (T = 0.88, year-unstable). Tiny opening ranges on DOT produce exactly those small gross moves. Filtering them out and checking stability by year targets both failure modes. (source: Structural Limits of OHLCV-Based Intraday Momentum Signals in MNQ Futures: A Systematic Falsification Study p.1)

Verification failed (Layer 4 — QA review) [class=hypothesis_mismatch]: - [edge_concern] The sandbox's 10 closed trades have negative net expectancy. Full-history evaluation should assess the literal short-direction rule after correctness is repaired; these losses do not determine this verdict. - [critical] should_enter() suppresses qualifying entries after end_time and clears both direction flags. The hypothesis explicitly preserves the source's literal entry and timed-liquidation rules. Source lines 85–103 check flat-only entry before timed liquidation, allowing a first qualifying breakout after the deadline to open and close in the same callback. Removing that round trip changes trades and costs; it is an additional rule absent from the hypothesis. (line 127) — fix: Preserve entry-before-liquidation sequencing, including a qualifying late entry's immediate close. Because FactoryStrategy dispatches either entry or exit per bar, implement the immediate close through an appropriate fill lifecycle hook without leaving the late position open until another bar. - [warning] When no eligible trade tick has been observed, calculate_signal() substitutes the bar close. The hypothesis calls for the latest observed trade price, so a bar-only replay can silently test a different signal input. Subscription flags are correctly enabled, but the supplied sandbox does not establish tick coverage. (line 107) — fix: Require an observed, point-in-time trade tick before evaluating entry conditions, and verify tick coverage for the tested window.