OPEN-SOURCE SCRIPT
Actualizado Streaming ML Probability Triple-Barrier, Conformal & Calibrated
Streaming-ML Probability — Triple-Barrier, Conformal & Calibrated
What it is
A self-training probability model that estimates P(up-barrier resolves before down-barrier) and — crucially — reports its own calibration. The differentiator is not the classifier; it is the labelling and the honesty layer wrapped around it. The pane shows a probability, an honest uncertainty band, a plain-language verdict, and a reliability diagram that tells you whether to believe any of it.
Why these components belong in ONE script (not a stack of indicators)
They are the stages of one honest prediction pipeline, each fixing a failure mode of a naive "ML" overlay:
Triple-barrier labels define what is predicted as a real, path-dependent outcome (which barrier is hit first) instead of an arbitrary "next bar up?", and resolve forward in time so training only ever sees confirmed results — no look-ahead.
Z-scored, bounded features keep every input finite so no single bar can blow up the online weights.
Logistic SGD + Lorentzian k-NN, fused in log-odds with a Kish decorrelation shrink — a linear model for the trend and a fat-tail-tolerant neighbour vote for non-linear structure, combined without double-counting the shared features.
Regime engine (efficiency + ADX + Hawkes) gives each market state its own calibration and shrinks the probability toward 0.5 where a regime is barely seen — the model defers to a coin flip where it has not learned.
Adaptive Conformal Inference turns the point estimate into a coverage-controlled band that holds under drift, so the uncertainty is honest rather than a fake point estimate.
Conviction + hard vetoes (regime, MTF, participation/CVD, calibration quality) stop the model acting on a number it cannot back up.
The calibration harness — reliability table, Brier score, and forward edge versus an unconditional base rate — is the whole point: a probability is only useful if 70% means 70%.
Labels say what to learn, features feed it, the fused classifier predicts, regime + conformal say how much to trust it, the vetoes gate it, and calibration proves whether any of it held. Remove a stage and the honesty breaks — that is why they ship as one engine.
How it works (mechanics)
Each confirmed bar opens a triple-barrier sample (unless the OU half-life says reversion is too slow to resolve in the horizon). When a sample resolves, the model takes one online gradient step on its own features and outcome, the resolved sample enters the k-NN memory, and the nonconformity score updates the conformal band via Gibbs-Candès ACI. The live probability is the log-odds fusion of the logistic output and the Lorentzian k-NN vote, regime-shrunk during warm-up. A directional signal fires only when the probability band clears the signal margin, conviction clears its floor, and no hard veto trips. Every resolved directional call is then scored against an unconditional same-horizon base rate (Hit% vs Base%, Wilson-bounded, regime- and recency-weighted).
Non-repaint: training only on resolved triple-barrier outcomes, signals on bar close, MTF requested with lookahead_off, no dynamic-length ta(). The live probability updates each bar — a current estimate from fixed historical training.
How to use
Read the VERDICT line first — ACT, STAND ASIDE (no calibrated edge), or WAIT, based on the live signal and the model's own calibration.
Check the reliability diagram / Brier. Points off the diagonal, or Brier ≥ 0.25 (worse than a coin flip), mean the probability is not trustworthy yet — the engine will tell you to stand aside.
A signal fires only on confluence: probability past the band-margin, conviction past the floor, no hard veto.
The conformal band is the honest uncertainty — a wide band means low confidence.
Everything here is descriptive, probabilistic context — never an instruction.
Note on honesty: on pure intraday noise (e.g. NIFTY 1-minute) this model will often show NO EDGE / VETO with Brier ≈ worse than a coin flip — and it says so plainly rather than inventing a signal. That is the intended behaviour.
Use on any market
The Data Source inputs (Close / High / Low / Volume) drive the features, the triple-barrier labels and the calibration, so the model runs on any series (standard candles, Heikin-Ashi, etc.) and any market. All thresholds are ATR-relative. Defaults are set for NIFTY index-futures intraday; change the source or lengths for other assets. Assets with no volume simply contribute nothing through the volume feature.
Originality
Most "machine-learning" indicators emit an uncalibrated score that is never checked against what actually happened. The contribution here is the closed, honest loop: real path-dependent labels, a fused-but-decorrelated classifier, regime-conditional shrinkage, a drift-robust conformal band, conviction/veto gating, and a built-in reliability + Brier + edge harness that can — and often does — tell you the model has no edge right now. It is built to be disprovable, which is the opposite of most signal scripts.
Credits
Triple-barrier labelling & meta-labelling — Marcos López de Prado
Logistic regression / SGD — classical statistics
Lorentzian (non-Euclidean) distance for k-NN — relativistic-distance concept
Conformal prediction — Vovk, Gammerman & Shafer; Adaptive Conformal Inference — Gibbs & Candès
Brier score — Glenn W. Brier
Wilson score interval — Edwin B. Wilson
Efficiency Ratio — Perry Kaufman · ADX / DMI — J. Welles Wilder
Hawkes self-exciting process — Alan G. Hawkes
Ornstein-Uhlenbeck / AR(1) half-life — Ornstein & Uhlenbeck
Effective-sample decorrelation — Leslie Kish
The pipeline assembly, the regime-conditional calibration and the conviction/veto layer are the author's original implementation.
Limitations (honest)
A well-calibrated probability is not an edge after costs. The forward stats are in-sample, close-to-close at a fixed horizon, with no costs, slippage or stops — a study aid, not a backtest. The model is small (six features, online weights); it warms up slowly and will stand aside often. Past behaviour does not assure future behaviour.
Disclaimer
Educational / informational study for chart analysis only. NOT financial advice, NOT a strategy, NOT a recommendation. It places no orders and guarantees no outcome. Markets carry risk; do your own research and manage your own risk. Paper-trade before risking real money.
What it is
A self-training probability model that estimates P(up-barrier resolves before down-barrier) and — crucially — reports its own calibration. The differentiator is not the classifier; it is the labelling and the honesty layer wrapped around it. The pane shows a probability, an honest uncertainty band, a plain-language verdict, and a reliability diagram that tells you whether to believe any of it.
Why these components belong in ONE script (not a stack of indicators)
They are the stages of one honest prediction pipeline, each fixing a failure mode of a naive "ML" overlay:
Triple-barrier labels define what is predicted as a real, path-dependent outcome (which barrier is hit first) instead of an arbitrary "next bar up?", and resolve forward in time so training only ever sees confirmed results — no look-ahead.
Z-scored, bounded features keep every input finite so no single bar can blow up the online weights.
Logistic SGD + Lorentzian k-NN, fused in log-odds with a Kish decorrelation shrink — a linear model for the trend and a fat-tail-tolerant neighbour vote for non-linear structure, combined without double-counting the shared features.
Regime engine (efficiency + ADX + Hawkes) gives each market state its own calibration and shrinks the probability toward 0.5 where a regime is barely seen — the model defers to a coin flip where it has not learned.
Adaptive Conformal Inference turns the point estimate into a coverage-controlled band that holds under drift, so the uncertainty is honest rather than a fake point estimate.
Conviction + hard vetoes (regime, MTF, participation/CVD, calibration quality) stop the model acting on a number it cannot back up.
The calibration harness — reliability table, Brier score, and forward edge versus an unconditional base rate — is the whole point: a probability is only useful if 70% means 70%.
Labels say what to learn, features feed it, the fused classifier predicts, regime + conformal say how much to trust it, the vetoes gate it, and calibration proves whether any of it held. Remove a stage and the honesty breaks — that is why they ship as one engine.
How it works (mechanics)
Each confirmed bar opens a triple-barrier sample (unless the OU half-life says reversion is too slow to resolve in the horizon). When a sample resolves, the model takes one online gradient step on its own features and outcome, the resolved sample enters the k-NN memory, and the nonconformity score updates the conformal band via Gibbs-Candès ACI. The live probability is the log-odds fusion of the logistic output and the Lorentzian k-NN vote, regime-shrunk during warm-up. A directional signal fires only when the probability band clears the signal margin, conviction clears its floor, and no hard veto trips. Every resolved directional call is then scored against an unconditional same-horizon base rate (Hit% vs Base%, Wilson-bounded, regime- and recency-weighted).
Non-repaint: training only on resolved triple-barrier outcomes, signals on bar close, MTF requested with lookahead_off, no dynamic-length ta(). The live probability updates each bar — a current estimate from fixed historical training.
How to use
Read the VERDICT line first — ACT, STAND ASIDE (no calibrated edge), or WAIT, based on the live signal and the model's own calibration.
Check the reliability diagram / Brier. Points off the diagonal, or Brier ≥ 0.25 (worse than a coin flip), mean the probability is not trustworthy yet — the engine will tell you to stand aside.
A signal fires only on confluence: probability past the band-margin, conviction past the floor, no hard veto.
The conformal band is the honest uncertainty — a wide band means low confidence.
Everything here is descriptive, probabilistic context — never an instruction.
Note on honesty: on pure intraday noise (e.g. NIFTY 1-minute) this model will often show NO EDGE / VETO with Brier ≈ worse than a coin flip — and it says so plainly rather than inventing a signal. That is the intended behaviour.
Use on any market
The Data Source inputs (Close / High / Low / Volume) drive the features, the triple-barrier labels and the calibration, so the model runs on any series (standard candles, Heikin-Ashi, etc.) and any market. All thresholds are ATR-relative. Defaults are set for NIFTY index-futures intraday; change the source or lengths for other assets. Assets with no volume simply contribute nothing through the volume feature.
Originality
Most "machine-learning" indicators emit an uncalibrated score that is never checked against what actually happened. The contribution here is the closed, honest loop: real path-dependent labels, a fused-but-decorrelated classifier, regime-conditional shrinkage, a drift-robust conformal band, conviction/veto gating, and a built-in reliability + Brier + edge harness that can — and often does — tell you the model has no edge right now. It is built to be disprovable, which is the opposite of most signal scripts.
Credits
Triple-barrier labelling & meta-labelling — Marcos López de Prado
Logistic regression / SGD — classical statistics
Lorentzian (non-Euclidean) distance for k-NN — relativistic-distance concept
Conformal prediction — Vovk, Gammerman & Shafer; Adaptive Conformal Inference — Gibbs & Candès
Brier score — Glenn W. Brier
Wilson score interval — Edwin B. Wilson
Efficiency Ratio — Perry Kaufman · ADX / DMI — J. Welles Wilder
Hawkes self-exciting process — Alan G. Hawkes
Ornstein-Uhlenbeck / AR(1) half-life — Ornstein & Uhlenbeck
Effective-sample decorrelation — Leslie Kish
The pipeline assembly, the regime-conditional calibration and the conviction/veto layer are the author's original implementation.
Limitations (honest)
A well-calibrated probability is not an edge after costs. The forward stats are in-sample, close-to-close at a fixed horizon, with no costs, slippage or stops — a study aid, not a backtest. The model is small (six features, online weights); it warms up slowly and will stand aside often. Past behaviour does not assure future behaviour.
Disclaimer
Educational / informational study for chart analysis only. NOT financial advice, NOT a strategy, NOT a recommendation. It places no orders and guarantees no outcome. Markets carry risk; do your own research and manage your own risk. Paper-trade before risking real money.
Notas de prensa
v2.0 — Calibration-honesty overhaul: Conformal-PID, isotonic recalibration, uniqueness-weighted training, deflated edgeConformal-PID band (Angelopoulos-Candès-Tibshirani 2023): ACI was only the integral term; a new proportional term reacts to the RECENT coverage error and tightens the band immediately under drift. Set pidKp = 0 to restore plain ACI.
Isotonic (pool-adjacent-violators) recalibration: the reported probability is passed through a monotone map fitted on THIS instrument's own reliability bins — "make 70% mean 70%." Fitted on past bars only, no feedback loop. Toggle useRecal; the panel marks the recalibrated read with ↺.
Sample-uniqueness-weighted training (López de Prado): overlapping triple-barrier labels aren't independent, so each SGD step is down-weighted by its concurrency — the effective sample size is honest.
Multiple-testing-deflated edge: the directional-edge ★ bar now rises with the number of edges tested (Bonferroni-style) instead of a flat 1.96, so a lucky edge isn't starred as real.
Exports: EXP_Pup now reports the RECALIBRATED probability; EXP_Praw (raw) and EXP_AlphaEff added.
Core engine, features, kNN memory, triple-barrier, MTF veto, theme and non-repaint behaviour unchanged; pidKp = 0 + useRecal = off reproduces v1 exactly.
Descriptive, self-calibrating probability tool; recalibrated on realised outcomes, not investment advice.
Script de código abierto
Fiel al espíritu de TradingView, el creador de este script lo ha convertido en código abierto, para que los traders puedan revisar y verificar su funcionalidad. ¡Enhorabuena al autor! Aunque puede utilizarlo de forma gratuita, recuerde que cualquier republicación del código está sujeta a nuestras Normas internas.
Exención de responsabilidad
La información y las publicaciones no constituyen, ni deben considerarse como, asesoramiento o recomendaciones financieras, de inversión, de trading u otro tipo, proporcionadas o respaldadas por TradingView. Obtenga más información en Condiciones de uso.
Script de código abierto
Fiel al espíritu de TradingView, el creador de este script lo ha convertido en código abierto, para que los traders puedan revisar y verificar su funcionalidad. ¡Enhorabuena al autor! Aunque puede utilizarlo de forma gratuita, recuerde que cualquier republicación del código está sujeta a nuestras Normas internas.
Exención de responsabilidad
La información y las publicaciones no constituyen, ni deben considerarse como, asesoramiento o recomendaciones financieras, de inversión, de trading u otro tipo, proporcionadas o respaldadas por TradingView. Obtenga más información en Condiciones de uso.