TT-Autotune — Strategy by ai_distribution_express
By ai_distribution_express
Performance Metrics
- Author: ai_distribution_express
- Symbol: TVC:USOIL
- Timeframe: 15 minutes
- Win Rate: 44.6%
- Profit Factor: 0.722
Description
## TT-Autotune — Machine-Optimized Lorentzian Classification with Live Regime SwitchingMost published strategies ship one set of parameters and hope the marketcooperates. TT-Autotune ships six — one per market regime — and each one wasbred, not guessed.TT-Autotune is a strategy fork of the well-known Machine Learning: LorentzianClassification by jdehorty (used under MPL-2.0, full credit to the originalauthor — the k-nearest-neighbors core with Lorentzian distance is his work).What we changed is everything around it: how the parameters are chosen, whenthey apply, and how the signals become orders.### The problem with the original defaultsLorentzian Classification is a brilliant classifier wrapped around ~30 tunableinputs: neighbor count, lookback depth, five feature slots (RSI / WT / CCI /ADX with two periods each), kernel regression settings, and four filters.The published defaults are one point in a 23-dimensional search space,calibrated by hand, for no particular symbol, on no particular timeframe.We measured that point. Across 50 crypto perpetual symbols on 15m, 1h, and 4h,the stock defaults produced a NEGATIVE median walk-forward Sortino (−0.15)net of fees. Not because the classifier is bad — because one parameter setcannot fit every market.### What TT-Autotune does differently1 — SERVER-BRED PARAMETERS (Sync Code)Every symbol × timeframe cell is optimized by TensorTrader's deterministicgenetic optimizer: populations of 32 candidate genomes evolved throughtournament selection, blend crossover, and decaying mutation (with fresh Sobolimmigrants every generation so the search never tunnel-visions), for a minimumof 12 and up to 128 generations. Fitness is not "biggest backtest profit" —it is walk-forward Sortino on net per-trade ROI after 5 bps fees + 2 bpsslippage, evaluated across 4 sequential out-of-sample folds on a pinned5,000-bar window, with a consistency penalty (mean fold Sortino minus half itsdispersion). A candidate needs 40+ trades and activity in every fold or it isdiscarded as statistically ineligible — no cherry-picked 6-trade miracles.The winning genome is serialized into a compact Sync Code you paste into oneinput field. Thirty parameters, injected at once.2 — LIVE REGIME SWITCHINGMarkets change character; your parameters should too. TT-Autotune embeds thesame six-state regime classifier that runs in TensorTrader's Python engine(ADX + linear-regression slope + ATR%, with confirmation-bar and minimum-holdhysteresis so it doesn't flip-flop): BULL_STRONG, BULL_WEAK, BEAR_STRONG,BEAR_WEAK, SIDEWAYS_QUIET, SIDEWAYS_CHOP. A packed Regime Sync Code carries aseparately-evolved champion for each regime, and the script hot-swaps theentire parameter set bar-by-bar as the live regime changes. An on-chart chippanel shows which regime is live, how long it has held, and which regimes havea deployed champion.3 — A FIXED NEIGHBOR POOLThe original script scans the OLDEST maxBarsBack bars of whatever historyTradingView happened to load — so the same settings give different signalsdepending on your chart's loaded history. TT-Autotune scans a sliding pool ofthe most RECENT maxBarsBack bars, making signals reproducible and matching thePython simulator the optimizer trains against, bar for bar.4 — A COMPLETE STRATEGY SHELLThis is a strategy, not an indicator: one alert covers long/short entries andexits, with optional pyramiding (up to 5 DCA legs across timeframes), ATR-basedstop-loss / take-profit brackets, session and date windows, and three signalmodes (DEFAULT ML entries/exits, KERNEL_RAW, KERNEL_SMOOTH kernel-fliptriggers).### The receiptsEvery optimization run evaluates the stock-default parameter set as a baselineon the exact same data, fees, folds, and eligibility rules as the candidates.Publication is fail-closed: a champion is only released if it is statisticallyeligible AND strictly beats the eligible default. From the most recentcompleted sweep (148 populations, 50 symbols × 15m / 1h / 4h, ~607,000walk-forward backtests):- Champion beat the stock default in 148 of 148 populations.- 136 of 148 cells flipped from a NEGATIVE default walk-forward Sortino to a positive champion score.- Median walk-forward fold Sortino: default −0.15 → champion +0.82.- Median consistency-adjusted improvement: +0.67 aggregate score per cell.- 825 of 888 per-regime panels produced an eligible regime champion; the rest honestly report a baseline fallback instead of faking a winner.Example (HBAR/USDT 4h): default settings scored 0.07 aggregate / +0.23 meanfold Sortino over 78 trades; the evolved champion scored 0.91 / +1.73 over 66trades — same window, same fees, same rules.These are walk-forward backtests net of realistic costs, not live results.Past performance never guarantees future returns. Trade small, trade paperfirst.### How to arm itWithout a Sync Code the script runs with the stock Lorentzian defaults — finefor exploring, but you are leaving the entire optimization layer on the table.To arm TT-Autotune with evolved parameters for your token:1. Create a free account at tensortrader.agent-swarm.net.2. Pick a symbol × timeframe cell from the optimized catalog (the free tier includes one active cell; paid plans scale with your TradingView alert quota).3. The TensorTrader browser extension delivers the packed Regime Sync Codes into the script's inputs and keeps your alert enrolled and heartbeat-fresh automatically — when a newer champion is bred for your cell, it rolls out to you.The codes are cell-specific on purpose: a champion evolved on HBAR 4h data hasno business trading DOGE 15m. The platform only arms the script for the marketit was actually trained on — that constraint is a feature, not a limitation.### Credits and licenseOriginal Lorentzian Classification logic © jdehorty, used under the MozillaPublic License 2.0. TT-Autotune's regime classifier, sync-code system,sliding neighbor pool, and strategy shell by TensorTrader.---## Publication notes (not part of the description)- TradingView House Rules require open-source forks of open-source scripts to credit the original and describe meaningful changes — the sections above do both explicitly.- Domain in the description is `tensortrader.agent-swarm.net` (the request contained a typo, "tnesortrader").- Figures verified 2026-08-07 against the 148 runs with `results.json` (admin-646 … admin-793 sweep, asof 2026-07-31). Regenerate with the aggregation snippet in the chat transcript before republishing if new runs have landed.