Model card: trade-level replay of the news veto
Browse files
README.md
CHANGED
|
@@ -119,6 +119,22 @@ its news filter is on par with the reference. Its 2× signal is less precise, so
|
|
| 119 |
with AI leverage locked off. Memory for `mlx-4bit`: about 5 GB idle; about 7.6 GB peak with 495 checks back to back. Speed on an Apple M5 Max: about 0.5 s
|
| 120 |
per request for either build.
|
| 121 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 122 |
### Extra questions (not trained, not used by the bot)
|
| 123 |
|
| 124 |
On test set B we also asked four questions Sentinel was not trained on. Zero-shot: event type (10 classes)
|
|
|
|
| 119 |
with AI leverage locked off. Memory for `mlx-4bit`: about 5 GB idle; about 7.6 GB peak with 495 checks back to back. Speed on an Apple M5 Max: about 0.5 s
|
| 120 |
per request for either build.
|
| 121 |
|
| 122 |
+
### What the veto did to the bot (replay over a year of trades)
|
| 123 |
+
|
| 124 |
+
Agreement with teacher labels is not the same as being useful for trading, so we also replayed the bot's news check
|
| 125 |
+
over a year of backtest entries (Binance perpetuals, 2025-10-01 to 2026-09-23, 20 coins, 1×, fees and funding
|
| 126 |
+
included), with archived headlines instead of the live feeds. Date-only timestamps count as known 24 hours later.
|
| 127 |
+
|
| 128 |
+
| | Entries | Avg. trade | Winners | 72 h after entry |
|
| 129 |
+
|---|---|---|---|---|
|
| 130 |
+
| Blocked by Sentinel (P ≥ 0.30) | 10 | −2.1% | 1 of 10 | −2.8% |
|
| 131 |
+
| Allowed | 606 | +1.0% | 38% | +0.9% |
|
| 132 |
+
|
| 133 |
+
Blocked entries did worse (−3.0 points per trade, one-sided permutation p ≈ 0.05), but the 10 blocks come from about
|
| 134 |
+
five events, so this is weak evidence. The portfolio result was unchanged (+61.4% with the news check, +61.5%
|
| 135 |
+
without; the strategy's stop-loss already limits these trades). Read: a guard against hack-type events, not a source
|
| 136 |
+
of returns. Code and details: [research/veto-eval](https://github.com/NovaeonStudio/novaeon-trading-ai/tree/main/research/veto-eval).
|
| 137 |
+
|
| 138 |
### Extra questions (not trained, not used by the bot)
|
| 139 |
|
| 140 |
On test set B we also asked four questions Sentinel was not trained on. Zero-shot: event type (10 classes)
|