Hypotheses

Faithful port of the open-source MIT strategy 'strategy-ml_lstm_multiple_features' by Quantiacs (https://github.com/quantiacs/strategy-ml_lstm_multiple_features, commit 51f5c6d). Credit goes to the original author. Any performance the source claims is unverified; our own backtest with real costs decides. The source asset NAS:AAPL maps to AAPL.USEQ on 1-DAY bars (split-adjusted, decades of history). The source backtest starts on 2006-01-01. It uses one instrument and OHLCV only, with no supplementary data. MODEL: two stacked LSTMCell layers (input_dim=3, hidden=64) followed by Linear(64->1). Hidden and cell states start at zero at the beginning of each sequence, and the model gives an output at every step. Gates follow standard PyTorch LSTMCell order i,f,g,o with both bias_ih and bias_hh. FEATURES per bar, in this order: [ln(close), ln(open), high]. The source does NOT log the high; this quirk is kept on purpose. Missing values are forward-filled, then back-filled, then filled with 1. TARGET for bar t: 1 if open[t+1] > open[t], else 0. In a training slice the last bar has no next open, so xr.where gives it label 0; this is kept. TRAINING: use the last train_period=100 bars up to and including the current bar as one batch of size 1. Loss is MSE over the 100 per-step outputs. The optimiser is L-BFGS with lr=0.08 and PyTorch defaults (max_iter=20, max_eval=25, tolerance_grad=1e-7, tolerance_change=1e-9, history_size=100, no line search), and exactly one optimiser.step call (epochs=1). Each retrain builds a fresh seed-42 model. Retraining happens every retrain_interval=360 bars; retrain_interval_after_submit is live-only and not used here. PREDICTION (predict_each_day=False, lookback_period=155): after each retrain, the model starts from zero state 155 bars before the retrain bar and runs causally until the next retrain. The output at bar d is the target weight for bar d. POSITION: w = clip(output, -1, 1). A positive w means long w x equity; a negative w means short |w| x equity. The position is rebalanced every daily bar to whole shares, decided at the close and executed at the next available price. The source has no stop or take-profit. Leverage is 1.0. DECLARED UNAVOIDABLE DEVIATION: PyTorch cannot be imported in the sandbox. The LSTMCell forward pass, the backprop-through-time gradient and PyTorch L-BFGS are therefore re-implemented in numpy. Initial weights are drawn U(-0.125, 0.125), PyTorch's default, from numpy default_rng(42) in PyTorch parameter order. Their distribution matches, but the exact torch RNG stream does not. Everything else is line-for-line.

A concrete long-short pairs strategy credited to Shinny (TqSdk), supplied under Apache-2.0. The original trades SHFE.ss2209 as Y and SHFE.ni2209 as X using 3600-second kline serials during 2022-07-04 through 2022-08-31. Preserve the supplied implementation exactly, including its unconventional Kalman equations, sizing and event timing. The inspected platform venue configuration does not support SHFE, and the brief establishes neither these contracts' historical data nor their margin and fee schedules. A faithful executable platform hypothesis is therefore infeasible with the established infrastructure. Substituting crypto instruments or completed-bar signals would produce a different experiment.

Faithful signal and order-rule translation of MultiTimeframeStrategy by 用Python的交易员, MIT licensed, at commit 6ef76981624bf55b2ea978f8587f74d633aafc72. Trade BCHUSDT.BYBIT long and short using completed 1-minute bars aggregated into calendar-aligned 5-minute and 15-minute bars. Source defaults remain rsi_signal=20, rsi_window=14, fast_window=5, slow_window=20, fixed_size=1. Initialize rsi_value, rsi_long, rsi_short, fast_ma, slow_ma and ma_trend to zero, then set rsi_long=70 and rsi_short=30 during initialization. Preserve two independent rolling ArrayManager-equivalent buffers of 100 bars and the source request for 10 days of historical initialization. The only platform sizing adaptation is that fixed_size=1 denotes one capital-relative allocation unit rather than one exchange contract. No performance claim is verified.

Faithful signal and order translation of AtrRsiStrategy by 用Python的交易员, MIT licensed. Source: https://github.com/vnpy/vnpy_ctastrategy/blob/6ef76981624bf55b2ea978f8587f74d633aafc72/vnpy_ctastrategy/strategies/atr_rsi_strategy.py. Preserve atr_length=22, atr_ma_length=10, rsi_length=5, rsi_entry=16, trailing_percent=0.8 and fixed_size=1. Use ETHUSDT.OKX perpetual with completed 1-MINUTE bars, matching the source's default tick-to-minute BarGenerator. No higher-timeframe aggregation, additional filters or parameter fitting. Platform sizing is the explicitly authorized adaptation: fixed_size=1 represents one capital-relative allocation, sized at entry to 10% of current equity in notional, rounded down to instrument quantity precision. Hold that quantity until exit; never pyramid. Config leverage=1.0. Starting equity is $100,000. Source initialization requests ten days of historical bars; maintain a rolling 100-bar ArrayManager-equivalent buffer and prohibit trading until it is initialized. Logging and event callbacks do not generate signals; order and stop-order callbacks are otherwise no-ops.

Translate SimpleBollinger by the Jesse community, whose source credits gabrielweich/jesse-strategies, under the supplied MIT licence. Trade DOTUSDT.BINANCE_SPOT on 1-hour bars, long-only at 1x. Preserve the source's entire-balance sizing, Bollinger breakout entry, Ichimoku entry filter and middle-band exit. No additional stop, target, volatility sizing, cooldown or trend-loss exit. Source: [pinned SimpleBollinger](https://github.com/jesse-ai/example-strategies/blob/7c91e0a37bf62165790120d730442e4f6eb00364/SimpleBollinger/__init__.py). Its performance claims are unverified.

Faithful signal translation of Chaos's Apache-2.0 TqSdk Aroon example, with explicitly adapted platform instrument and sizing. Source defaults are SYMBOL=SHFE.au2306, POSITION_SIZE=30 contracts, START_DATE=2023-02-20, END_DATE=2023-05-05, daily bars (86400 seconds), AROON_PERIOD=10, AROON_UPPER_THRESHOLD=75, AROON_LOWER_THRESHOLD=25 and STOP_LOSS_PCT=1.2. The platform experiment trades DOTUSDT.BINANCE using the same indicator, decision ordering and defaults, with capital-relative sizing instead of 30 gold contracts. This tests a crypto port, not reproduction of the gold-contract performance. Crucially, the source evaluates when the newest daily bar's datetime changes: at creation of the forming bar, not its eventual close. Build that opening snapshot causally from trade ticks and completed daily history. Do not substitute completed-day Aroon signals. [TqSdk documents this event timing](https://doc.shinnytech.com/tqsdk/latest/usage/backtest.html).

Faithful translation of QuantRocket LLC's Apache-2.0 UpMinusDown signal and allocation rules, with explicitly declared platform instruments and costs. Source: https://github.com/quantrocket-codeload/moonshot-intro/blob/8717e87e279e3615bdd4c7787ce746b92d0713d7/intro_moonshot/umd.py. Preserve MOMENTUM_WINDOW=252, RANKING_PERIOD_GAP=22, TOP_N_PCT=50 and REBALANCE_INTERVAL='M'. Use synchronized USEQ daily session bars. The base UpMinusDown class does not specify a universe; this experiment declares a fixed ten-stock platform universe and does not claim to reproduce the demo's unspecified usstock-free membership. UpMinusDownDemo additionally sets CODE='umd-demo', DB='usstock-free-1d', UNIVERSES='usstock-free', TOP_N_PCT=50 and COMMISSION_CLASS=USStockCommission; that commission class specifies $0.005 per share. Our experiment uses USEQ's zero commissions with modeled spread and impact instead. No optimization, additional filter, stop-loss or signal-strength sizing is permitted.

Faithful fixed-rule translation of the Jesse community strategy, credited to the supplied author and its referenced gabrielweich/jesse-strategies repository. Trade ATOMUSDT.BINANCE_SPOT using 4-HOUR bars, long-only, without leverage. The source does not specify an instrument or interval; these are platform experiment choices. Preserve its full-balance sizing and channel exit. Performance is unverified. Source: [pinned Donchian strategy](https://github.com/jesse-ai/example-strategies/blob/7c91e0a37bf62165790120d730442e4f6eb00364/Donchian/__init__.py).

Faithful translation of TradingView_RSI by Jesse community, MIT licensed. Trade ATOMUSDT.BINANCE_SPOT on completed 4-HOUR bars, long-only, without leverage. The source specifies neither instrument nor interval; these are platform experiment choices. Preserve RSI period 5, stop_loss multiplier 0.95, take_profit multiplier 1.1, and xparam 75. The declared RSI search range 10–30 excludes its default 5; retain 5 rather than silently correcting it. All source defaults are frozen, with no optimization or added filters. Starting capital is $100,000. Performance is unverified.

A source-inspired adaptation of Guosen Securities' market-strength-conditioned month-start effect, published by BigQuant 宽邦. Trade SPY.USEQ using 1-DAY bars: remain long in strong markets; in weaker markets, go long before calendar day 15 and short from day 15 onward. The original report concerns Chinese indices. This experiment tests transferability to a liquid US index ETF; it does not reproduce the report or adopt its claimed returns. Because the full PDF's RPS definition could not be verified, use the explicitly registered causal proxy below.

One fixed-parameter experiment adapting the source's 55-day breakout, four-unit pyramiding and independent half-ATR Whipsaw stops to ATOM, NEAR and UNI perpetuals. Trade both directions using completed UTC daily channels and 4-hour execution bars. Each stopped unit can re-enter at its original entry level while its breakout campaign remains active. Source performance claims are unverified. This is an adaptation, not an exact reproduction: execution is close-confirmed, risk is reduced, and re-entry frequency is bounded.

Test a disclosed adaptation of the source's percentage recovery from rolling extremes and midpoint exits on ATOMUSD.HYPERLIQUID, using completed 4-HOUR bars. This is not statistical regression. The prose permits simultaneous long and short conditions; resolve that ambiguity using threshold crossings and mutually exclusive midpoint conditions. Preserve percentage thresholds without adding RSI, moving-average trend filters or cross-sectional ranking. Source performance claims are unverified.

Test the source's bullish moving-average alignment and pullback-entry mechanism on three fixed crypto spot instruments using daily bars. Preserve SMA periods 5/10/20/60 and symmetric 10% profit/loss thresholds. Explicit adaptations: replace historical A-share selection with a predeclared crypto basket, replace 1000-share sizing with equity-relative sizing, and evaluate a closing-price pullback at the completed daily bar instead of claiming execution at an opening price already in the past. This is an adaptation, not a faithful reproduction. The source's performance claims provide no evidence for this experiment.

Cross-venue, long-short event trade. TRADED legs: ADAUSDT.OKX, AVAXUSDT.OKX, LINKUSDT.OKX, SOLUSDT.OKX (OKX USDT perpetual swaps, 4H bars, history 2020-03 to 2021-01 onward). SIGNAL legs: the same coins on Binance SPOT (ADAUSDT/AVAXUSDT/LINKUSDT/SOLUSDT.BINANCE_SPOT, 4H bars), which are never traded. Idea: a large move that is carried by spot buying or selling is unlevered demand or supply, i.e. holders adding or dumping inventory. It tends to keep going, because spot flow is slower and more persistent than levered perp flow. A move of the same size carried mostly by perp volume is often leverage churn that unwinds. We measure who drove the bar with the log ratio of Binance spot base volume to OKX perp base volume (both base-unit, so the units cancel), z-scored against its own trailing 180-bar (30-day) distribution. This removes structural venue-share drift. We trade only the spot-led half of the shocks and in their direction. Per coin: shock = |log return of the 4H OKX close| / trailing 180-bar std of 4H returns >= shock_sigma. Spot-led = spot share z >= spot_share_z. Enter in the direction of the shock bar at the next bar, hold a fixed hold_bars (default 12 = 48h), with a protective ATR stop. Each coin has an independent slot; position = 15% of equity per coin, at most 4 concurrent (60% gross), 1x, no leverage. Flat roughly 85-90% of the time. No supplementary feed is needed (L189). Venue choice: OKX is 2.2% of experiments and cross-venue is 7.4%, both under quota, and long-short is 17.4%. BTC/ETH are deliberately excluded (over-concentrated). These four are the OKX swaps with 5+ years of history in the local catalog (DOGE/XRP OKX history is only days). Distinct from the abandoned 'Perp-Led Selloff Absorption' (which used premium index, Binance USD-M only, long-only fade) and from same-coin cross-venue dislocation arb (fee-dead): here the second venue is used only for information, so the trade has one leg per position and pays a single ~0.10% round trip. Prior evidence: the price-shock/volume-confirmation family has survivors in live_stage (LINK Bybit shock absorption, ETH Bybit thrust continuation). This adds a participant-composition filter rather than another volume threshold.

Test whether an unusually large relative-price gap compresses after correlated trading strengthens and the outperformer begins losing relative strength. Use synchronized 4-HOUR bars for LTC, LINK, AVAX and DOT Hyperliquid perpetuals; all four exact bar directories were confirmed locally. Hold one dollar-matched long–short pair at a time. Source inspiration: https://bigquant.com/wiki/doc/50Zn3u05OH — its interaction between asset clustering and relative valuation motivates this experiment. This is a new crypto adaptation, not a reproduction: relative price is not fundamental valuation, and the article's equity findings do not establish crypto profitability. Use only the four instruments' overlapping history, with no synthetic extensions or symbol substitutions.

Test whether a failed breakout in Binance XRP spot, followed by a confirming completed bar, predicts a substantial directional reversal in the Hyperliquid XRP perpetual. Use synchronized 4-HOUR bars; trade only XRPUSD.HYPERLIQUID, with XRPUSDT.BINANCE_SPOT as an untraded signal source. Inspiration: the Spring and Upthrust concepts described at https://www.prorealcode.com/prorealtime-indicators/wyckoff-event-detection/; this is a new falsifiable trading specification, not a reproduction of its indicator. Both exact bar directories exist locally. Evaluate their actual overlapping history, without assuming the directory proves sufficient coverage.

Trade a dollar-balanced LTC/BCH perpetual pair on synchronized 4-hour bars. Test whether a large relative-price move accompanied by unusually concentrated trading activity partially reverses after the spread first turns back. Both exact catalog directories exist, with filenames indicating overlapping history from January 2024 to October 2026; continuity still needs validation. This is a price-convergence experiment, not a funding-income strategy. Implement through FactoryStrategy extra-instrument hooks, with one paired position at a time.

Test whether an unusually large, high-volume asset-specific daily move that fails to persist reverses over the following 2–7 days. Trade BCH, ETC, LINK and ATOM perpetuals against an equal-dollar basket of the other three instruments. This is an event-driven failed-auction hypothesis, not a low-volatility ranking factor or a claim of cointegration. Daily OHLCV supplies synchronized returns, volume and closing-location confirmation. Only one four-leg cohort may be active. Positions use capital-relative sizing, 0.5% equity risk per cohort and at most 0.60x gross exposure. The proposed universe comes from the supplied brief; aligned catalog coverage must be checked before implementation.

Calendar-structure, market-neutral event trade on OKX USDT perpetual swaps. OKX is 2.1% of the corpus, and the design is long-short and cross-sectional; both buckets are under quota. The weekend range is the high and low of the 48 1H bars that close in (Sat 00:00, Mon 00:00] UTC. Weekend liquidity is thin and largely retail, so the weekend low is a crowded stop and liquidation magnet. In the first 48h of the week (Mon 00:00 to Wed 00:00 UTC), the first 1H close BELOW the weekend low takes out those stops; weekday institutional flow then absorbs them and the alt recovers relative to the market. In-house prestudy on OKX 1H bars, 2020-2026, 8 pre-declared coins (ADA, DOT, BCH, ATOM, LINK, ETC, DOGE, TRX), 1,316 events, gross of costs. Buying the alt at that close and holding 96h returned +2.05% raw. The BTC return over the same window was +1.01%, so the BTC-hedged spread returned +1.04% mean. By year (hedged): 2020 -1.30%, 2021 +3.97%, 2022 +0.84%, 2023 +0.93%, 2024 +1.03%, 2025 +0.76%, 2026 YTD -0.07%. By coin: 7 of 8 positive (TRX -0.10%, DOGE +3.83% is the outlier and inflates the mean). Win rate is only 47%, so the payoff is right-skewed. The mirror trade, shorting an upside break of the weekend high, measured about -0.4% to -0.5% across hold lengths and is deliberately EXCLUDED. The naive breakout-FOLLOW version on the same data lost about 1% per trade on 6 of 8 coins, which is the same finding seen from the other side: early-week breaks of the weekend range fail. Hold-length sensitivity on the unhedged long side: 72h +1.0%, 96h +2.0%, 120h +2.1%. That curve is flat-ish rather than a spike, but 48h was about 0, so the clock is a real degree of freedom; it is fixed at 96h to avoid tuning it. Mechanism differs from the abandoned DOGE Bybit-OKX dislocation (same-coin, cross-venue spread) and from generic RSI/Bollinger reversion. The trigger is a calendar-anchored liquidity level, not a z-score.

This is a faithful translation of 'Oil Money Trading backtest' by je-suis-tm (quant-trading repo, Apache-2.0, commit 611b73f2c3f577ac5b28aaa19ac8c43d3236c7a5). All credit for the strategy goes to the author. The code is a tradeable strategy (signal_generation produces long/short signals). The remaining functions (oil_money, plot, profit, and the portfolio mark-to-market) are scaffolding and visualization. INSTRUMENT MAPPING, the one unavoidable substitution: the original regresses NOKJPY (y, traded) on Brent crude (x, regressor). The platform has no FX or Brent feed. The closest tradeable analogue with the same economic structure (an oil-linked asset regressed on the oil price, trading the oil-linked asset) and decades of history is y = XOP.USEQ (S&P Oil & Gas E&P ETF, traded, primary) and x = USO.USEQ (crude oil fund, regressor/signal only, never traded). Both have daily bars from mid-2006, about 5,100 overlapping sessions. Venue: USEQ, Reg-T margin, shorts allowed, zero commission with roughly 0.02-0.05% round-trip spread and impact modeled. Timeframe: 1-DAY bars, as in the original daily data. ALGORITHM, line for line from signal_generation (defaults holding_threshold=10, stop=0.5, rsquared_threshold=0.7, train_len=50). State: holding in {-1,0,1}, trained=False, counter=0, frozen coefficients (a,b), sigma, entry_price. The loop starts at bar index train_len, so the first 50 aligned bars are warm-up. On each aligned daily bar i (XOP and USO closes for the same session both present): (A) If holding != 0: (A1) if counter > holding_threshold, exit (signal = -holding), set holding=0, trained=False, counter=0, and do NOTHING else this bar (continue, no refit and no entry). (A2) Otherwise, if |close_y[i] - entry_price| >= stop, exit the same way and continue. (A3) Otherwise counter += 1. Note that the counter starts at 0 on the entry bar and is incremented after the checks, so the time exit fires on the 12th bar after the entry bar. (B) If holding == 0: (B1) if not trained, fit OLS with a constant, y ~ a + b*x, on the PREVIOUS 50 bars [i-50, i), excluding bar i. If R-squared > rsquared_threshold (strict), set trained=True, freeze a and b, and set sigma = population std (ddof=0) of the in-sample residuals. If R-squared fails, stay untrained and retry the fit on the next bar with the window rolled forward. (B2) If trained (this includes the bar where the model was just fitted), forecast_i = a + b*close_x[i] using the CURRENT bar's USO close and the FROZEN coefficients. If close_y[i] > forecast_i + 2*sigma, go LONG (holding=1, entry_price=close_y[i]). Else if close_y[i] < forecast_i - 2*sigma, go SHORT (holding=-1, entry_price=close_y[i]). Note the polarity: the source goes LONG when the traded asset is ABOVE the upper band and SHORT when BELOW the lower band. This is a residual-breakout / follow direction, not a fade. Translate it exactly and do not 'correct' it to mean reversion. While trained and flat, the model is never refit. The same frozen a, b and sigma are applied to each new x until a trade happens and then exits (exit sets trained=False and forces a fresh fit on the next bar). The 1-sigma bands and the band-collapsing after entry only affect the plot and do not change trading logic, so omit them. PARAMETER TRANSLATION: stop=0.5 is in absolute NOKJPY points. NOKJPY traded around 12-14 over the author's sample, so 0.5 points is about 3.7% of price. The port expresses it scale-free as stop_fraction=0.037 of entry_price (|close/entry_price - 1| >= 0.037), because 0.5 dollars on a ~$40-150 ETF would be a different rule. All other defaults are kept verbatim. SIZING: the original buys a fixed share count capital0//max(close over the WHOLE series), which is look-ahead and is replaced. The port sizes each entry at position_fraction=0.95 of current account equity in whole XOP shares (1x, about the original's all-in notional), for longs and shorts alike. EXECUTION: the source fills at the signal bar's close. On the platform, signals are evaluated on the daily close and filled by market order at the next session open. The stop and time checks reference entry_price = the signal-bar close, as in the source (df[y][signals!=0].iloc[-1]). PRE-STUDY (research lead's pandas replay of the exact loop on catalog closes, 2006-2026, close-to-close fills, no costs; unverified): 144 trades, mean +0.38%/trade, win rate 56%, profit factor 1.16. Robustness check with XLE as y: 137 trades, +0.22%/trade, PF 1.10. The edge is thin, so the platform backtest decides.

Translate the supplied tradeable CtaTemplate strategy, credited to 用Python的交易员, without correcting its unusual ladder or fill callbacks. Preserve source defaults exactly: entry_window=20, exit_window=10, atr_window=20, fixed_size=1. Use LTCUSD.HYPERLIQUID perpetuals, long and short, with LTCUSD.HYPERLIQUID-4-HOUR-LAST-EXTERNAL bars. Four-hour input is an explicit platform adaptation: the original tick-fed BarGenerator normally builds minute bars, while Hyperliquid minute history is insufficient for substantial validation. Capital-relative sizing is the other explicit adaptation: fixed_size=1 represents one campaign-frozen risk unit rather than literally one LTC. No added indicator, regime filter, profit target, time stop or funding signal. Preserve initialization of all channel, entry, stop and ATR variables to zero; source initialization requests load_bar(20), meaning historical days rather than a 20-bar warm-up. Logging and put_event are lifecycle/display operations. Source on_order and on_stop_order callbacks contain no trading logic. Performance is unverified.