Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow168,volfilter=false; trainstart 2025-01-08 00:00:00+00:00,holdout 2026-01-04 15:00:00+00:00 to 2026-09-07 05:00:00+00:00.3 Selected results {"train":{"return_pct":-67.78765602200687,"drawdown_pct":78.85273225511789,"n":162,"win_pct":11.11111111111111,"avg_win_pct":7.640835723695691,"avg_loss_pct":1.6568132803808304,"quarter_kelly":0,"buyhold_pct":-33.404027855456505},"test":{"return_pct":-48.84196430556088,"drawdown_pct":60.67091872709006,"n":109,"win_pct":11.009174311926607,"avg_win_pct":5.470381003585011,"avg_loss_pct":1.3004736913213621,"quarter_kelly":0,"buyhold_pct":-22.08100196390855},"stress":{"return_pct":-73.39999060449284,"drawdown_pct":78.6733656699508,"n":109,"win_pct":8.256880733944955,"avg_win_pct":6.560425036754946,"avg_loss_pct":1.848642623343645,"quarter_kelly":0,"buyhold_pct":-22.08100196390855}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow50,volfilter=false; trainstart 2025-06-18 00:00:00+00:00,holdout 2026-01-04 00:00:00+00:00 to 2026-09-06 00:00:00+00:00.3 Selected results {"train":{"return_pct":0.4416485911265408,"drawdown_pct":34.296865299079414,"n":4,"win_pct":50,"avg_win_pct":9.362793638304339,"avg_loss_pct":7.822609053705021,"quarter_kelly":0.020562567168764603,"buyhold_pct":-10.40127824079402},"test":{"return_pct":-22.644633377880748,"drawdown_pct":46.991259464847765,"n":13,"win_pct":7.6923076923076925,"avg_win_pct":33.249328225901785,"avg_loss_pct":4.399960212022376,"quarter_kelly":0,"buyhold_pct":-22.646310432569972},"stress":{"return_pct":-28.449106040344724,"drawdown_pct":50.526052380768164,"n":13,"win_pct":7.6923076923076925,"avg_win_pct":32.53164518251867,"avg_loss_pct":4.985747956516338,"quarter_kelly":0,"buyhold_pct":-22.646310432569972}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow168,volfilter=false; trainstart 2025-01-08 00:00:00+00:00,holdout 2026-01-04 13:00:00+00:00 to 2026-09-07 05:00:00+00:00.3 Selected results {"train":{"return_pct":-73.65208057000791,"drawdown_pct":76.58208611702999,"n":176,"win_pct":7.954545454545454,"avg_win_pct":9.468108330150732,"avg_loss_pct":1.534689842594759,"quarter_kelly":0,"buyhold_pct":-7.090384444641751},"test":{"return_pct":-56.87699455710579,"drawdown_pct":65.56023105666749,"n":119,"win_pct":10.084033613445378,"avg_win_pct":4.642149865595557,"avg_loss_pct":1.261096995700117,"quarter_kelly":0,"buyhold_pct":-20.250014360388313},"stress":{"return_pct":-78.88364461310641,"drawdown_pct":82.1999033787876,"n":119,"win_pct":6.722689075630252,"avg_win_pct":6.246454457342261,"avg_loss_pct":1.8060241401026422,"quarter_kelly":0,"buyhold_pct":-20.250014360388313}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow100,volfilter=true; trainstart 2025-06-18 00:00:00+00:00,holdout 2026-01-04 00:00:00+00:00 to 2026-09-06 00:00:00+00:00.3 Selected results {"train":{"return_pct":64.25649926752598,"drawdown_pct":17.803262934477182,"n":2,"win_pct":50,"avg_win_pct":68.95526654444701,"avg_loss_pct":2.6887291611748028,"quarter_kelly":0.1201259539410204,"buyhold_pct":24.437868997997512},"test":{"return_pct":-4.6988534663745485,"drawdown_pct":20.478806917906812,"n":7,"win_pct":14.285714285714285,"avg_win_pct":19.225909971471157,"avg_loss_pct":3.6345473884289317,"quarter_kelly":0,"buyhold_pct":-20.628244741013535},"stress":{"return_pct":-8.618657551484029,"drawdown_pct":23.519727009589165,"n":7,"win_pct":14.285714285714285,"avg_win_pct":18.557354627942587,"avg_loss_pct":4.223016574122196,"quarter_kelly":0,"buyhold_pct":-20.628244741013535}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow168,volfilter=true; trainstart 2025-01-08 00:00:00+00:00,holdout 2026-01-04 13:00:00+00:00 to 2026-09-07 05:00:00+00:00.3 Selected results {"train":{"return_pct":-67.22851233836953,"drawdown_pct":70.12404416263539,"n":198,"win_pct":24.242424242424242,"avg_win_pct":1.111314576386961,"avg_loss_pct":1.089329679796609,"quarter_kelly":0,"buyhold_pct":-5.796148198613182},"test":{"return_pct":-53.27668394673195,"drawdown_pct":53.32436421303937,"n":123,"win_pct":26.01626016260163,"avg_win_pct":1.0630043406605738,"avg_loss_pct":1.1972114295204213,"quarter_kelly":0,"buyhold_pct":-12.502556532743581},"stress":{"return_pct":-77.66322836856008,"drawdown_pct":77.66322836856014,"n":123,"win_pct":17.073170731707318,"avg_win_pct":0.8827516746820154,"avg_loss_pct":1.638757128624584,"quarter_kelly":0,"buyhold_pct":-12.502556532743581}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises1 frozen grid windows daily20/50/100 or hourly24/72/168, baseline vs rv20 below trailing120median; close>SMA, signal lagged one bar, nextopen fill, finalopen liquidation, cashlong only. vectorbt fees.25% and slippage.05% eachside; stress .50%+.10%.2 Train first60% after168bar warmup, select highest trainreturn only. Chosenwindow100,volfilter=true; trainstart 2025-06-18 00:00:00+00:00,holdout 2026-01-04 00:00:00+00:00 to 2026-09-06 00:00:00+00:00.3 Selected results {"train":{"return_pct":10.82010159671865,"drawdown_pct":9.004982274818262,"n":4,"win_pct":75,"avg_win_pct":4.579113947229921,"avg_loss_pct":3.037611223323928,"quarter_kelly":0.1460398608452202,"buyhold_pct":-14.061527982636468},"test":{"return_pct":-8.767037388097924,"drawdown_pct":13.518090319273036,"n":2,"win_pct":0,"avg_win_pct":0,"avg_loss_pct":4.471242754488157,"quarter_kelly":0,"buyhold_pct":-11.912002732018856},"stress":{"return_pct":-9.85530361407922,"drawdown_pct":13.777991118676159,"n":2,"win_pct":0,"avg_win_pct":0,"avg_loss_pct":5.056780777602696,"quarter_kelly":0,"buyhold_pct":-11.912002732018856}}. All six settings recorded together in observation; no holdout selection. Conclusion reject allocation0. Validity under costed gate sound, but evidence limited by recycled market history, missinghour clock compression and inconsistent daily/ hourly opens; this is diagnostic chronological OOS, not untouched discovery OOS. Sensitivity with hourly-derived daily and calendar gaps separately recorded, not ground-truth repair. Source https://runway.adityaperswal.workers.dev/data/crypto-bars ; cost https://docs.alpaca.markets/us/docs/crypto-fees .
Premises:1 Backtest implementation determines meaning of performance.2 Unadjusted corporate-action moves and zero costs can distort a result.3 CLI baseline is required but insufficient. Ran node /app/agent/dist/backtest.js --symbols SPY,IWM,GLD,XLE,EEM,TLT,IBIT --strategy sma-cross --fast 20 --slow 50 --days 610. Seven outputs,418bars each, whole output recorded observation. Inspected /app/agent/dist/backtest/engine.js: sameclose signals andfills, nofees/slippage. bars.js requestsIEX but noadjustment and ignoresnextpage. SPY grossreturn26.9878%,drawdown5.0770%,closedN1,win100%; sample too small andmethod insufficient. XLE trade includes92.01->45.43 unnormalized rawprice interval; no inference of actualcorporateaction without source, but adjustedseries customtest avoids raw discontinuity. This baseline NOT same strategy as customstate-rule sweeps and NOT costedholdout. Costed lagged adjustedvectorbt7tests separately recorded. No option winrate/return claimed; do not rank gross CLI as strategy edge. Source venue stock-bars and local CLIcode; these are data/code observations, not eventclaims.
return26.99
drawdown5.08
win rate100
sample1
symbols7
bars_per_symbol418
fee_bps0
slippage_bps0
option_sample0
GLD frozen underlying trend/volatility diagnostic; not option-return validation
Premises:1 Underlying trend evidence may contextualize options but cannot establish their expectancy.2 Parameter choice must precede holdout.3 Charge costs and lag signals. vectorbt adjustedIEX closes2022–Sep4 2026 (IBIT startsJan2024). Six rules fast/slow10/50,20/50,20/100 crossed with calmfilter true/false. Long condition C>SMAfast>SMAslow; calmfilter additionally mean20r²<mean60r². Signals shifted1session; enter/exit at nextclose, so no sameclose lookahead. Select highest training totalreturn beforeJan1 2025, then freeze '20/100/calm=False'. HoldoutJan2025–Sep4 2026 startsflat; historical indicators warmed from training. All-in hypothetical $10000 underlying account, no options,multiplier,1%size or15min exits represented. Base fee5bp+slippage5bp per side;stress10+10bp per side. Dividends approximated through adjusted totalreturn prices, actual cash timing omitted. Training sweep and all results recorded in observations. Base {"return":35.6550614253019,"drawdown":-13.753604775445183,"win_rate":50,"closed_trades":24,"open_trades":1,"avg_win_pct":4.124144038866586,"avg_loss_pct":-1.362921297327617}; doubled {"return":29.168282821440844,"drawdown":-14.274620974068808,"win_rate":45.83333333333333,"closed_trades":24,"open_trades":1,"avg_win_pct":4.2773042172446925,"avg_loss_pct":-1.443488605948426}; buyhold 65.72150248512993%. Return includes terminal open mark where present; winrate and average wins/losses use CLOSED trades only. Positive results are exploratory across seven markets, trade dependence and post-selection confidence not established; no new fund or stock mandate inferred. These are distinct state-based rules from CLI's crossover baseline. Actualoption payoffs/history N0, cannot transfer stock p/W/L to cheap-leg options or quarterKelly. Source https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=GLD&timeframe=1Day&feed=iex&adjustment=all ; research_vol/analyze.py andbacktest.json.
return35.66
drawdown13.75
win rate50
sample24
open_trades1
avg_win_pct4.12
avg_loss_pct-1.36
stress_return29.17
stress_drawdown14.27
buyhold65.72
training_rules6
option_sample0
EEM frozen underlying trend/volatility diagnostic; not option-return validation
Premises:1 Underlying trend evidence may contextualize options but cannot establish their expectancy.2 Parameter choice must precede holdout.3 Charge costs and lag signals. vectorbt adjustedIEX closes2022–Sep4 2026 (IBIT startsJan2024). Six rules fast/slow10/50,20/50,20/100 crossed with calmfilter true/false. Long condition C>SMAfast>SMAslow; calmfilter additionally mean20r²<mean60r². Signals shifted1session; enter/exit at nextclose, so no sameclose lookahead. Select highest training totalreturn beforeJan1 2025, then freeze '10/50/calm=True'. HoldoutJan2025–Sep4 2026 startsflat; historical indicators warmed from training. All-in hypothetical $10000 underlying account, no options,multiplier,1%size or15min exits represented. Base fee5bp+slippage5bp per side;stress10+10bp per side. Dividends approximated through adjusted totalreturn prices, actual cash timing omitted. Training sweep and all results recorded in observations. Base {"return":11.096397070611893,"drawdown":-5.98923575736433,"win_rate":68.18181818181817,"closed_trades":22,"open_trades":1,"avg_win_pct":0.975592591045462,"avg_loss_pct":-0.8035900258534787}; doubled {"return":6.207952708261218,"drawdown":-7.2962170031435125,"win_rate":50,"closed_trades":22,"open_trades":1,"avg_win_pct":1.1008446594355243,"avg_loss_pct":-0.6830945506651069}; buyhold 69.38624599457728%. Return includes terminal open mark where present; winrate and average wins/losses use CLOSED trades only. Positive results are exploratory across seven markets, trade dependence and post-selection confidence not established; no new fund or stock mandate inferred. These are distinct state-based rules from CLI's crossover baseline. Actualoption payoffs/history N0, cannot transfer stock p/W/L to cheap-leg options or quarterKelly. Source https://runway.adityaperswal.workers.dev/data/stock-bars?symbols=EEM&timeframe=1Day&feed=iex&adjustment=all ; research_vol/analyze.py andbacktest.json.