Per-Asset Performance
Crypto
Bitcoin
Directional Accuracy
58.6%
BUY / SELL correct direction
Overall 3-Class Accuracy
44.8%
BUY / SELL / WAIT · random = 33.3%
Crypto
Ethereum
Directional Accuracy
58.9%
BUY / SELL correct direction
Overall 3-Class Accuracy
44.8%
BUY / SELL / WAIT · random = 33.3%
Commodity
Gold
Directional Accuracy
62.4%
BUY / SELL correct direction
Overall 3-Class Accuracy
50.2%
BUY / SELL / WAIT · random = 33.3%
Commodity
Silver
Directional Accuracy
62.0%
BUY / SELL correct direction
Overall 3-Class Accuracy
43.8%
BUY / SELL / WAIT · random = 33.3%
Energy
Oil (WTI)
Directional Accuracy
60.2%
BUY / SELL correct direction
Overall 3-Class Accuracy
51.0%
BUY / SELL / WAIT · random = 33.3%
Energy
Brent Oil
Directional Accuracy
60.5%
BUY / SELL correct direction
Overall 3-Class Accuracy
50.9%
BUY / SELL / WAIT · random = 33.3%
Index
S&P 500
Directional Accuracy
62.1%
BUY / SELL correct direction
Overall 3-Class Accuracy
50.4%
BUY / SELL / WAIT · random = 33.3%
Equity
NVIDIA
Directional Accuracy
62.5%
BUY / SELL correct direction
Overall 3-Class Accuracy
51.1%
BUY / SELL / WAIT · random = 33.3%
Equity
Apple
Directional Accuracy
63.4%
BUY / SELL correct direction
Overall 3-Class Accuracy
50.3%
BUY / SELL / WAIT · random = 33.3%
Index
Dow Jones
Directional Accuracy
61.3%
BUY / SELL correct direction
Overall 3-Class Accuracy
50.2%
BUY / SELL / WAIT · random = 33.3%
Why two accuracy numbers?
The model outputs three classes: BUY, SELL, or WAIT. Overall accuracy measures all three — including WAIT signals, which are deliberately issued when conviction is low. Directional accuracy only measures the BUY/SELL calls against a 50% coin-flip baseline. That is the number that matters for trading: when the model commits to a direction, how often is it right?
Conviction Calibration — The Higher the Confidence, the More Accurate
| Confidence |
Bitcoin |
Ethereum |
Gold |
Silver |
Oil (WTI) |
Brent |
Dow Jones |
S&P 500 |
NVIDIA |
Apple |
Rating |
| 45 – 55% |
46.4% (59.9% dir) |
45.8% (59.8% dir) |
46.9% (60.6% dir) |
48.3% (60.7% dir) |
48.2% (58.1% dir) |
48.8% (59.3% dir) |
47.6% (60.0% dir) |
47.9% (59.7% dir) |
47.4% (59.4% dir) |
46.8% (61.6% dir) |
Moderate |
| 55 – 65% |
54.5% (66.8% dir) |
57.5% (69.7% dir) |
62.5% (70.1% dir) |
61.5% (68.8% dir) |
60.8% (67.9% dir) |
63.3% (69.7% dir) |
64.2% (70.5% dir) |
63.6% (70.7% dir) |
61.6% (69.7% dir) |
64.2% (72.7% dir) |
High |
| 65 – 75% |
63.6% (72.2% dir) |
71.1% (79.0% dir) |
81.2% (81.7% dir) |
78.8% (78.0% dir) |
75.1% (78.2% dir) |
81.7% (83.2% dir) |
80.9% (82.3% dir) |
78.8% (79.6% dir) |
76.5% (77.9% dir) |
84.3% (83.0% dir) |
High |
| 75%+ |
84.0% (81.2% dir) |
85.1% (87.8% dir) |
91.3% (88.0% dir) |
87.1% (78.0% dir) |
90.1% (86.2% dir) |
93.5% (90.7% dir) |
91.3% (89.5% dir) |
86.8% (86.0% dir) |
89.5% (85.3% dir) |
95.4% (90.0% dir) |
Peak |
What this means: Conviction is monotonically calibrated — accuracy rises consistently as the model's confidence increases. At 75%+ conviction, BTC signals have been correct 84.0% of the time across 430 historical instances, and ETH signals 85.1% correct across 322 instances. These high-conviction signals are rare by design: the model holds back unless the evidence is strong. Users can see the current conviction level on every signal in the app.
Signal Quality — Correct Calls Move More
Why the quality gap matters: Even when the model is wrong about direction, the magnitude of the move tends to be smaller. When it is right, the move is larger. This asymmetry — correct calls catching bigger moves, incorrect calls occurring on smaller noise — is the structural trading edge beyond raw accuracy alone.
How the Model Works
Signal Type
3-class classification — BUY, SELL, or WAIT. The WAIT class is a deliberate output, not a fallback. The model withholds a directional call when evidence is insufficient, reducing noise and false positives.
Label Generation
Maximum Favorable Excursion (MFE) scaled by Average True Range. Labels reflect whether a real, volatility-adjusted move occurred in the next 6 bars — not just whether the next candle closed up or down.
Feature Engineering
100+ technical features per bar — momentum indicators, volatility measures, volume signals, trend structure, candlestick patterns, Ichimoku components, and multi-timeframe lags. All features are strictly backward-looking.
Dimensionality Reduction
Return-weighted PCA — a variation of standard PCA where principal components are selected based on their correlation with forward price returns, not just variance explained. This biases the feature space toward market-relevant structure.
Model Architecture
Soft-voting ensemble of gradient-boosted trees and a random forest. Each model produces class probabilities; the ensemble averages them. The final confidence score is the probability of the winning class — a real calibrated number, not a heuristic.
Validation Protocol
Strict chronological split — training on earliest data, testing on most recent. No shuffling. 5-fold time-series cross-validation with each fold advancing forward in time. No data leakage between folds.
Comparative Summary
| Asset |
Directional Accuracy |
Overall Accuracy |
Uplift vs Coin Flip |
Uplift vs Random (3-class) |
Rating |
|
|
62.4% |
50.2% |
+12.4pp |
+16.8pp |
High |
|
|
58.6% |
44.8% |
+8.6pp |
+11.5pp |
High |
|
|
58.9% |
44.8% |
+8.9pp |
+11.5pp |
High |
|
|
62.0% |
43.8% |
+12.0pp |
+10.5pp |
High |
|
|
60.2% |
51.0% |
+10.2pp |
+17.6pp |
High |
|
|
60.5% |
50.9% |
+10.5pp |
+17.6pp |
High |
|
|
61.3% |
50.2% |
+11.3pp |
+16.8pp |
High |
|
|
62.1% |
50.4% |
+12.1pp |
+17.1pp |
High |
|
|
62.5% |
51.1% |
+12.5pp |
+17.8pp |
High |
|
|
63.4% |
50.3% |
+13.4pp |
+17.0pp |
High |
Past model performance does not guarantee future results. Markets are non-stationary, so model accuracy may vary across different market regimes. Enodara provides market intelligence for informational and educational purposes only and does not constitute financial advice. Enodara is not a registered broker-dealer, investment adviser, or financial advisor, and does not manage client funds or execute trades. All figures will be updated following each model retrain.