ML Predictions: A Deep Dive Into the Probability Models Behind CoreNova

Frameworks

Most "AI stock prediction" tools sell magic. CoreNova's ML layer sells one specific thing: a probability that price will move directionally over the next N bars, produced by three ensemble models, decomposed into auditable feature contributions. This is the engineering deep-dive, including what we DON'T do — like publish a single win-rate number or call this a "signal."

Of all the words abused in retail trading marketing, "AI" is probably the worst — followed closely by "ML" and "prediction." YouTube is full of services promising "92% accurate AI signals" with no explanation of what model is running, what data it's trained on, or what "accuracy" even means in a probabilistic context. The result: traders pay subscription fees for black boxes that produce signals they can't audit, evaluate, or improve on. We built CoreNova's ML layer to do the opposite. We tell you exactly what models run, exactly what features they consume, exactly how they produce the probability, and exactly what that probability is supposed to mean (and not mean).

This is a deep dive into the ML predictions framework — one of the 9 frameworks feeding the Cross-Tool Consensus and the AI Trade Strategist. You'll learn the three ensemble models we actually run (XGBoost, LightGBM, Random Forest), the ~50 engineered features they consume (and why feature engineering matters more than algorithm choice), the per-asset/per-timeframe model architecture, the SHAP feature attribution that makes stock predictions auditable, the honest calibration story (probabilities are calibrated, with a stock-side guard that rejects calibration whenever it would degrade the fit), and how ML probability fits into the broader 9-framework synthesis. This is the operator's guide to what "ML predictions" actually means at the implementation level.

3 — Ensemble models
~50 — Engineered features
Per-TF — Model granularity
1 of 7 — Methodology votes

Four-tier ML prediction pipeline. Input: raw OHLCV candles for the asset + timeframe. Feature Engineering: ~50 engineered features grouped by category (returns, MAs, momentum, volatility, volume, structure). Models: three ensemble models run in parallel — XGBoost as primary, LightGBM as speed-optimized fallback, Random Forest as comparison baseline. Output: a single probability number (P of upside move over next N bars) plus top feature contributions via SHAP attribution (stocks). The probability becomes one of the framework votes in the Cross-Tool Consensus.

Probabilities, Not Signals — The Core Thesis

The single most important thing to understand about CoreNova's ML layer is what it outputs. It does NOT output a directive ("BUY!"). It does NOT output an arrow on a chart. It does NOT output a price target. It outputs ONE number: the probability that price will move directionally over the next N bars (where N depends on the timeframe — typically 5 bars on 4H, 10 bars on 1H, etc.).

Why this distinction matters: a probability is a meaningfully different kind of output than a signal. A signal pretends to know what will happen on this specific trade. A probability tells you what would happen over many trades with similar conditions. A 65% probability means "in 100 setups like this, about 65 of them historically went up." Any individual trade can go either way — but if you take enough of them, you converge to the long-run rate. This is the only honest framing for predictive models in probabilistic markets.

The single insight that defines our ML layer The ML probability is one framework vote among nine — not the answer. When ML probability is 70% bullish but Wyckoff says distribution and Elliott counts a complete impulse, the trade isn't obvious. The probability informs direction; the chartist methodologies inform structure; the AI Trade Strategist synthesizes. Treating the ML score as a standalone "signal" is the failure mode we deliberately designed around.

The Three Ensemble Models

Every ML prediction in CoreNova passes through three ensemble models in parallel. We don't pick one and commit; we run all three because they fail in different ways and consensus across them is meaningful.

XGBoost: Gradient-boosted decision trees. The workhorse — strongest out-of-sample accuracy on our feature set. Handles non-linear feature interactions natively. Used as the primary probability source on every analysis. Battle-tested across nearly every Kaggle tabular competition for the last decade.
LightGBM: Histogram-based gradient boosting from Microsoft Research. Materially faster than XGBoost on large feature sets (10-30× faster inference in practice). Used when latency matters or as a secondary check on XGBoost output. Similar accuracy profile, different training algorithm.
Random Forest: Bagged decision trees. Less accurate than gradient-boosted models on average, but catches non-linear feature interactions in different ways. Run as a baseline — when Random Forest disagrees with the boosted models, that's a flag that the feature space may be unusual.

Why three models and not one Each algorithm has different bias-variance tradeoffs and different failure modes. XGBoost can overfit on noisy regimes. LightGBM can be sensitive to outlier features. Random Forest underfits subtle interactions. When all three agree, you have a robust signal across paradigms. When they disagree, the probability is treated as lower-confidence — which the AI Trade Strategist incorporates. This is much harder to do with a single model.

Per-Asset, Per-Timeframe Model Granularity

We don't ship one mega-model trained on "everything." That doesn't work for financial data — different assets behave differently, and different timeframes have fundamentally different microstructure. Instead, every asset class × timeframe combination has its own trained model.

Asset Class	Timeframes	Model File Pattern
Stocks (e.g., AAPL, MSFT, NVDA)	5m, 15m, 30m, 1h, 4h, 1d	AAPL5m.pkl, AAPL1h.pkl, AAPL_1d.pkl, etc.
Crypto (e.g., BTC, ETH, SOL)	5m, 15m, 1h, 4h, 1d	BTC5m.pkl, BTC4h.pkl, ETH_1d.pkl, etc.
Universal indices (sector / market regime)	1h, 4h, 1d	Macro context models — supplement asset-specific predictions

Why this matters: a 5-minute BTC chart and a 1-day AAPL chart have nothing in common at the feature level. The volatility regimes differ. The volume dynamics differ. The relevant patterns differ. A single model trained across both would be a worst-of-both-worlds compromise. By training a separate model per combination, each one specializes — and we can update individual models as regimes shift without retraining everything.

Feature Engineering — Why Algorithm Choice Is the Easy Part

The dirty secret of machine learning in finance is that model algorithm matters far less than feature engineering. XGBoost, LightGBM, Random Forest, and a dozen other algorithms all converge to similar accuracy IF the features are good. What separates predictive systems is what the model is allowed to see. Raw OHLCV alone produces noisy, low-information models. Engineered features that encode trader-relevant context produce signals that actually have predictive value.

CoreNova's ML layer trains on ~50 engineered features — not raw OHLCV. The features fall into 6 categories:

Returns & log returns: 1, 3, 5, 10, 20-bar simple returns. 1-bar and 5-bar log returns. These are the model's view of recent price velocity at multiple time horizons. Returns are non-stationary; log returns help with the heavy-tailed distribution.
Moving averages & crossovers: SMA(5/10/20/50) and EMA(5/12/26). Plus relationships: price vs SMA20, price vs SMA50, SMA5 vs SMA20, SMA20 vs SMA50. The relationships matter as much as the levels — they encode trend structure.
Momentum indicators: RSI(14) and RSI(7), MACD line / signal / histogram, Stochastic %K and %D, Rate-of-Change at multiple windows, Momentum(10). The classic momentum triad plus extensions.
Volatility: ATR(14), ATR ratio, Bollinger Bands upper/lower/width/position, realized volatility at 5/10/20-bar windows. Volatility regime is one of the strongest predictors of regime-specific behavior — and one of the most informative feature categories.
Volume & order flow: Volume SMA(10/20), volume ratio, On-Balance Volume slope, volume-price trend. Plus on crypto specifically: order-flow imbalances derived from real exchange order book data — institutional positioning signals not available on stocks.
Structure & patterns: Higher-highs-in-last-3, lower-lows-in-last-3, consecutive up/down bars, distance from 20-bar high/low, candle anatomy (upper/lower shadow, body range). These encode price-action structure the model can use without us hand-coding chart patterns.

Each feature is computed deterministically from the OHLCV — no AI generates them — and then passed into the three ensemble models as input. The model learns which features matter and how to combine them; the engineering decides what the model gets to see.

Reading an ML Prediction Output

Here's what a real ML prediction output looks like in the CoreNova UI. The probability isn't just a number — it's decomposed into the feature contributions that drove it, so you can audit the reasoning.

Sample ML prediction output for AAPL on the 4H timeframe with a 5-bar horizon. The probability gauge at the top shows 64% upside probability (model marker at the 64% mark of the 0-100 scale). Below, the top 5 feature contributions are shown SHAP-style — RSI(14) oversold recovery is the strongest positive contributor (+0.34), MACD histogram turning positive adds +0.21, volume confirmation adds +0.18, the SMA crossover adds +0.14, while distance-from-20-bar-low leans slightly bearish (-0.06). The interpretation box translates the math into a plain-English summary. The disclaimer at the bottom is critical: this is ONE methodology vote of 7, not a buy signal.

Calibration — What "Probability" Should Mean (and Whether We Achieve It)

Here's the part that most "AI signal" services don't talk about: probability calibration. A well-calibrated model should produce probabilities that match real-world outcomes. If the model says 65% probability of upward move, then over many such predictions, ~65% should actually go up. If 80% of "65% predictions" go up, the model is under-confident at that bucket; if only 50% do, it's over-confident.

What probability calibration means. The green diagonal line is the "perfect calibration" reference — model probability equals actual win rate. The blue dots are example data: each represents hundreds of historical predictions at that probability bucket, plotted against how often they actually went up. A model whose dots cluster near the diagonal is well-calibrated. A model whose dots systematically sit below the diagonal is over-confident; above means under-confident. The honest caveat: CoreNova's XGBoost models are NOT formally calibrated with isotonic regression or Platt scaling — we surface raw probabilities and tell you so, rather than hiding behind a calibration layer that may itself be miscalibrated.

Honest disclosure — we don't formally calibrate CoreNova's ML models produce raw probabilities from XGBoost / LightGBM / Random Forest, without applying isotonic regression or Platt scaling to enforce calibration. The probabilities may be slightly over- or under-confident relative to ground truth — and we don't pretend otherwise. The feature attribution helps you contextualize: if the probability is 70% but it's leaning on one anomalous feature, you can downweight your conviction accordingly. We chose honesty over a false calibration layer.

Retraining Cadence — Adapting to Regime Shifts

Markets are not stationary. A model trained on 2023's regime won't necessarily generalize to 2026's regime. Volatility expands and contracts. Correlations between assets shift. Order-flow patterns change as new participants enter. A static model deployed once and never retrained will degrade — sometimes within months.

Our solution: rolling-window retraining. Each per-asset, per-timeframe model is retrained on a rolling window of recent data, with the older data progressively dropped out. The exact window size and retraining cadence depends on the timeframe (5m models retrain more frequently than 1d models because they have more training data per unit time). The effect: models adapt to regime shifts rather than getting stale.

Rolling-window training data: only the most recent N bars are used; older bars age out. The model never sees a regime it has no information about.
Per-timeframe retraining cadence: 5m and 15m models retrain more often (more data flowing in per unit time); 4h and 1d models retrain on slower cycles.
Out-of-sample validation built in: when a retrained model performs worse on held-out validation data than the previous version, we don't deploy it — the previous model stays active until the new one outperforms.
Regime-shift alerts: when validation performance degrades suddenly (faster than rolling adaptation can handle), the system flags a potential regime shift and the per-framework confidence weighting adjusts.

ML Probability in the 9-Framework Consensus

The ML probability score is ONE of the framework votes in CoreNova's Cross-Tool Consensus. It is never the only input to a trade decision. Here's how it integrates:

As one vote: ML probability sits alongside Wyckoff phase, Elliott Wave count, Fibonacci levels, Ichimoku Cloud signal, Gann angles, and technical-indicator consensus. All 7 contribute to the consensus.
Multi-paradigm confirmation: When ML + Wyckoff + Elliott + Ichimoku all align bullish, you have a setup confirmed across data-driven (ML) AND chartist (Wyckoff / Elliott / Ichimoku) paradigms. This is the strongest signal type.
Disagreement surfacing: When ML probability is 70% bullish but Wyckoff says distribution phase and Elliott counts a complete impulse, the AI Trade Strategist explicitly flags the disagreement rather than picking a side. Honest synthesis beats false consensus.
Confidence weighting: When the three ensemble models (XGBoost / LightGBM / Random Forest) disagree with each other, the ML score is weighted lower in the consensus calculation. Inter-model disagreement is itself information.
Asset-class-specific context: On stocks, ML probability sits alongside options-chain signals (Put/Call, IV rank, gamma). On crypto, it sits alongside order book depth, Bitcoin Network Health, and Crypto Fear & Greed. The full asset-class picture is incorporated.

What the ML Layer Does NOT Do

Equally important is what we deliberately don't claim or do:

It does not publish a single "win rate" or "accuracy" number. Markets are non-stationary; win rate is a function of regime, timeframe, and trade selection. Any single accuracy number is misleading. We publish methodology — you compute your own performance.
It does not predict exact prices. The output is a directional probability, not a price target. Targets come from chart structure (Fibonacci extensions, prior swing highs, etc.) — see our Risk Management guide.
It does not replace risk management. A high ML probability does not authorize a larger position. Risk is still bounded by structure-derived stops and the 1% rule.
It does not execute trades or connect to your broker. Same as everything else in CoreNova — we're an analysis platform. The ML probability goes into your decision; you place the order.
It does not hide its features. Every prediction comes with the top 5 feature contributions visible. If the model is leaning on something you find suspicious, you can downweight your conviction.
It does not claim to be calibrated when it isn't. We disclose openly that we don't formally calibrate with isotonic regression or Platt scaling. Raw probabilities surfaced.

Plan Availability

ML predictions are included on every paid CoreNova plan:

Plan	Price	Asset Class	ML Models Included
Stock Analysis Pro	$59/month	Stocks only	Stock-specific XGBoost / LightGBM / Random Forest models across all supported timeframes
Crypto Analysis Pro	$59/month	Crypto only	Crypto-specific models (with order-flow features) across all supported timeframes
Complete Bundle ★	$99/month (7-day trial)	Both stocks AND crypto	All ML models for both asset classes — unified consensus across the full toolkit

See the ML probability + top-5 feature attribution on real charts, alongside the 8 other framework verdicts and the AI Trade Strategist synthesis. Bundle 7-day trial covers both stocks AND crypto. Start Free Trial

Five Mistakes Retail Traders Make With ML Predictions

Treating the probability as a buy signal. "70% bullish" doesn't mean "buy now." It means setups like this historically went up about 70% of the time over the relevant horizon. Any individual trade can still go either way. The probability informs direction; structure (and the 8 other methodologies) inform execution.
Ignoring the feature attribution. A 70% probability driven by 5 well-balanced features is a different signal than a 70% probability where one anomalous feature contributes +0.6 and everything else is neutral. Always scroll to the top-5 features — if one feature is doing all the work, downweight your conviction.
Looking for "win rates" or "accuracy numbers." These metrics are misleading without regime-specific context. A model with 65% accuracy in a bull market may have 45% accuracy in a chop regime. We deliberately don't publish a single number; the probability + feature attribution + multi-framework context is the honest version.
Trusting calibration without verification. Even formally calibrated models become miscalibrated as regimes shift. Our models aren't formally calibrated and we say so. Either way, never trust a probability blindly — combine it with chart-structure context from the other methodologies.
Letting ML override chartist methodologies. When ML probability says 70% bullish but Wyckoff clearly shows a distribution phase, the AI Trade Strategist surfaces the disagreement — and you should respect it. Multi-paradigm disagreement is information. Don't pick the answer you like.

Frequently Asked Questions

What is CoreNova's ML predictions framework?

CoreNova's ML predictions framework is one of the 9 frameworks feeding the Cross-Tool Consensus. It outputs a single number: the probability that price will move directionally over the next N bars (typically 5 bars on 4H, 10 bars on 1H, varying by timeframe). The probability is produced by three ensemble models — XGBoost as primary, LightGBM as a speed-optimized fallback, Random Forest as a comparison baseline — trained on ~50 engineered features per asset × timeframe combination. On stocks, every prediction comes with SHAP feature attributions so you can audit which features drove the score. The probability is then incorporated into the multi-framework consensus that powers the AI Trade Strategist.

Which ML models does CoreNova actually use?

Three ensemble models, run in parallel on every prediction: (1) XGBoost — the primary model, gradient-boosted decision trees, strongest out-of-sample accuracy on our feature set; (2) LightGBM — Microsoft Research's histogram-based gradient boosting, faster inference, used as a speed-optimized fallback or secondary check; (3) Random Forest — bagged decision trees, less accurate on average but catches non-linear interactions differently, used as a comparison baseline. When all three agree, the probability is treated as higher-confidence; when they disagree, the score is downweighted in consensus. Each asset × timeframe combination has its own trained model file (BTC4h.pkl, AAPL1d.pkl, etc.) — we don't ship one mega-model trained across everything.

What features do the ML models actually train on?

Approximately 50 engineered features, NOT raw OHLCV. The feature set is organized into 6 categories: (1) Returns at multiple windows (1, 3, 5, 10, 20-bar returns plus log returns); (2) Moving averages and their relationships (SMA 5/10/20/50, EMA 5/12/26, plus crossover relationships); (3) Momentum indicators (RSI 14 and 7, MACD line/signal/histogram, Stochastic %K/%D, Rate-of-Change at multiple windows); (4) Volatility (ATR 14 and its ratio, Bollinger Bands upper/lower/width/position, realized volatility at 5/10/20 windows); (5) Volume and order flow (volume SMA, volume ratio, OBV slope, volume-price trend; on crypto specifically: real exchange order-flow imbalances); (6) Structure and patterns (higher-highs/lower-lows count, consecutive up/down bars, distance from 20-bar high/low, candle anatomy). Feature engineering matters more than algorithm choice — getting features right is most of the work.

Are the ML models calibrated?

Honest answer: not formally. We don't apply isotonic regression or Platt scaling to enforce calibration on the raw XGBoost / LightGBM / Random Forest probability outputs. We surface the raw probabilities and disclose this openly. The reasoning: a poorly-applied calibration layer can be worse than no calibration at all (it can mask regime shifts that would otherwise show up as miscalibration). Instead, we provide top-5 feature attributions for every prediction so you can contextualize the score — if the probability is leaning heavily on one anomalous feature, you can downweight your conviction. We chose honest disclosure over a calibration layer that may itself be miscalibrated.

How often are the ML models retrained?

Each model is retrained on a rolling window so it adapts to regime shifts rather than becoming stale. The exact cadence varies by timeframe — 5m and 15m models retrain more frequently because they accumulate more training data per unit calendar time; 4h and 1d models retrain on slower cycles. The retraining workflow includes out-of-sample validation: when a freshly retrained model performs worse on held-out validation data than the currently-deployed model, we don't promote it. The previous model stays active until the new one outperforms. This prevents accidental model regressions from being silently deployed.

How is this different from 'AI signal services' on YouTube?

Three structural differences. First, our output is a probability, not a directive — we don't say "BUY!" because that's not how probabilistic models work over many trades. Second, every prediction is decomposed into top-5 feature contributions you can audit — no black box. Third, the ML score is ONE framework vote among nine in our consensus engine, not the whole answer. When ML disagrees with Wyckoff or Elliott, the disagreement is surfaced rather than hidden. Additionally: we don't publish single "win rate" or "accuracy" numbers because those metrics are misleading without regime context. We publish methodology — you compute your own performance on your trades.

Does the ML model predict price targets or stop levels?

No. The ML model outputs directional probability only. Price targets and stop levels come from CHART STRUCTURE — not from the ML output. Targets are derived from Fibonacci extensions, prior swing highs/lows, chart-pattern measured moves, or other structure-based projections. Stops are placed at structural invalidation levels (Wyckoff swing lows, Fib boundaries, Ichimoku cloud edges, Elliott wave invalidation points). This separation is intentional: structure-derived levels reflect where the actual market participants placed orders, not arbitrary numbers from a model. See our Risk Management deep dive for the full discussion of structure-based stops vs percentage-based stops.

Read “ML Predictions: A Deep Dive Into the Probability Models Behind CoreNova” on CoreNova Analytics