6093eac7bf
Wave 64-65 cleanup: Proto regeneration and build system updates from Tonic 0.12→0.14 upgrade Files updated: - Cargo.lock: Dependency resolution for Tonic 0.14.2 - All build.rs: Updated for tonic-prost-build - Proto files: Regenerated with tonic-prost 0.14 - Examples/tests: Updated for new gRPC API 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
391 lines
14 KiB
Rust
391 lines
14 KiB
Rust
//! Numerical equivalence tests for ML model accuracy validation
|
|
//!
|
|
//! Critical for HFT production deployment to ensure our Rust implementations
|
|
//! produce numerically equivalent results to reference Python implementations.
|
|
|
|
use std::collections::HashMap;
|
|
use anyhow::{Result, Context};
|
|
use serde::{Deserialize, Serialize};
|
|
use tokio::time::{Duration, timeout};
|
|
|
|
use crate::{
|
|
MLModel, ModelType, Features, ModelPrediction,
|
|
mamba::Mamba2SSM,
|
|
dqn::RainbowDQN,
|
|
tlob::TLOBTransformer,
|
|
tft::TemporalFusionTransformer,
|
|
};
|
|
|
|
/// Tolerance levels for numerical comparison in HFT context
|
|
#[derive(Debug, Clone)]
|
|
pub struct NumericalTolerance {
|
|
/// Absolute tolerance for exact comparisons
|
|
pub absolute: f64,
|
|
/// Relative tolerance for proportional comparisons
|
|
pub relative: f64,
|
|
/// Maximum acceptable difference for HFT decisions
|
|
pub decision_threshold: f64,
|
|
}
|
|
|
|
impl Default for NumericalTolerance {
|
|
fn default() -> Self {
|
|
Self {
|
|
absolute: 1e-10, // 10 decimal places precision
|
|
relative: 1e-8, // 8 significant figures
|
|
decision_threshold: 1e-6, // 1 millionth for trading decisions
|
|
}
|
|
}
|
|
}
|
|
|
|
/// Test case for numerical equivalence validation
|
|
#[derive(Debug, Clone, Serialize, Deserialize)]
|
|
pub struct NumericalTestCase {
|
|
pub name: String,
|
|
pub model_type: ModelType,
|
|
pub input_features: Features,
|
|
pub expected_output: ModelPrediction,
|
|
pub tolerance: NumericalTolerance,
|
|
pub description: String,
|
|
}
|
|
|
|
/// Results of numerical equivalence testing
|
|
#[derive(Debug, Clone)]
|
|
pub struct NumericalTestResult {
|
|
pub test_name: String,
|
|
pub passed: bool,
|
|
pub rust_output: ModelPrediction,
|
|
pub python_reference: ModelPrediction,
|
|
pub absolute_error: f64,
|
|
pub relative_error: f64,
|
|
pub latency_ns: u64,
|
|
pub error_message: Option<String>,
|
|
}
|
|
|
|
/// Comprehensive numerical validation framework
|
|
pub struct NumericalValidator {
|
|
test_cases: Vec<NumericalTestCase>,
|
|
models: HashMap<ModelType, Box<dyn MLModel>>,
|
|
python_bridge: Option<PythonModelBridge>,
|
|
}
|
|
|
|
impl NumericalValidator {
|
|
/// Create new validator with comprehensive test suite
|
|
pub fn new() -> Self {
|
|
Self {
|
|
test_cases: Self::create_standard_test_cases(),
|
|
models: HashMap::new(),
|
|
python_bridge: None,
|
|
}
|
|
}
|
|
|
|
/// Register Rust model for testing
|
|
pub fn register_model(&mut self, model_type: ModelType, model: Box<dyn MLModel>) {
|
|
self.models.insert(model_type, model);
|
|
}
|
|
|
|
/// Initialize Python bridge for reference comparisons
|
|
pub fn with_python_bridge(mut self, bridge: PythonModelBridge) -> Self {
|
|
self.python_bridge = Some(bridge);
|
|
self
|
|
}
|
|
|
|
/// Run complete numerical validation suite
|
|
pub async fn validate_all_models(&mut self) -> Result<Vec<NumericalTestResult>> {
|
|
let mut results = Vec::new();
|
|
|
|
for test_case in &self.test_cases {
|
|
match self.validate_single_test(test_case).await {
|
|
Ok(result) => results.push(result),
|
|
Err(e) => {
|
|
results.push(NumericalTestResult {
|
|
test_name: test_case.name.clone(),
|
|
passed: false,
|
|
rust_output: ModelPrediction::default(),
|
|
python_reference: ModelPrediction::default(),
|
|
absolute_error: f64::INFINITY,
|
|
relative_error: f64::INFINITY,
|
|
latency_ns: 0,
|
|
error_message: Some(e.to_string()),
|
|
});
|
|
}
|
|
}
|
|
}
|
|
|
|
Ok(results)
|
|
}
|
|
|
|
/// Validate single test case with timing
|
|
async fn validate_single_test(&mut self, test_case: &NumericalTestCase) -> Result<NumericalTestResult> {
|
|
let model = self.models.get(&test_case.model_type)
|
|
.context("Model not registered for testing")?;
|
|
|
|
// Time the Rust inference
|
|
let start = std::time::Instant::now();
|
|
let rust_output = timeout(
|
|
Duration::from_millis(100), // 100ms timeout for HFT
|
|
model.predict(&test_case.input_features)
|
|
)
|
|
.await
|
|
.context("Inference timeout")?
|
|
.context("Inference failed")?;
|
|
let latency_ns = start.elapsed().as_nanos() as u64;
|
|
|
|
// Get Python reference if available
|
|
let python_reference = if let Some(ref bridge) = self.python_bridge {
|
|
bridge.predict(&test_case.model_type, &test_case.input_features).await?
|
|
} else {
|
|
test_case.expected_output.clone()
|
|
};
|
|
|
|
// Calculate numerical differences
|
|
let absolute_error = self.calculate_absolute_error(&rust_output, &python_reference)?;
|
|
let relative_error = self.calculate_relative_error(&rust_output, &python_reference)?;
|
|
|
|
// Determine if test passed
|
|
let passed = absolute_error <= test_case.tolerance.absolute &&
|
|
relative_error <= test_case.tolerance.relative &&
|
|
absolute_error <= test_case.tolerance.decision_threshold;
|
|
|
|
Ok(NumericalTestResult {
|
|
test_name: test_case.name.clone(),
|
|
passed,
|
|
rust_output,
|
|
python_reference,
|
|
absolute_error,
|
|
relative_error,
|
|
latency_ns,
|
|
error_message: None,
|
|
})
|
|
}
|
|
|
|
/// Calculate absolute error between predictions
|
|
fn calculate_absolute_error(&self, rust: &ModelPrediction, python: &ModelPrediction) -> Result<f64> {
|
|
match (rust, python) {
|
|
(ModelPrediction::Price(r), ModelPrediction::Price(p)) => {
|
|
Ok((r.value - p.value).abs())
|
|
},
|
|
(ModelPrediction::Direction(r), ModelPrediction::Direction(p)) => {
|
|
Ok(if r.direction == p.direction { 0.0 } else { 1.0 })
|
|
},
|
|
(ModelPrediction::Portfolio(r), ModelPrediction::Portfolio(p)) => {
|
|
let mut total_error = 0.0;
|
|
for (symbol, rust_pos) in &r.positions {
|
|
if let Some(python_pos) = p.positions.get(symbol) {
|
|
total_error += (rust_pos.quantity - python_pos.quantity).abs();
|
|
}
|
|
}
|
|
Ok(total_error)
|
|
},
|
|
_ => Ok(f64::INFINITY), // Type mismatch
|
|
}
|
|
}
|
|
|
|
/// Calculate relative error between predictions
|
|
fn calculate_relative_error(&self, rust: &ModelPrediction, python: &ModelPrediction) -> Result<f64> {
|
|
let absolute_error = self.calculate_absolute_error(rust, python)?;
|
|
|
|
match python {
|
|
ModelPrediction::Price(p) => {
|
|
if p.value.abs() < 1e-15 {
|
|
Ok(absolute_error)
|
|
} else {
|
|
Ok(absolute_error / p.value.abs())
|
|
}
|
|
},
|
|
ModelPrediction::Direction(_) => Ok(absolute_error),
|
|
ModelPrediction::Portfolio(p) => {
|
|
let total_value: f64 = p.positions.values()
|
|
.map(|pos| pos.quantity.abs())
|
|
.sum();
|
|
if total_value < 1e-15 {
|
|
Ok(absolute_error)
|
|
} else {
|
|
Ok(absolute_error / total_value)
|
|
}
|
|
},
|
|
}
|
|
}
|
|
|
|
/// Create comprehensive test cases for all models
|
|
fn create_standard_test_cases() -> Vec<NumericalTestCase> {
|
|
vec![
|
|
// MAMBA-2 SSM Tests
|
|
NumericalTestCase {
|
|
name: "mamba_sequence_modeling".to_string(),
|
|
model_type: ModelType::MAMBA,
|
|
input_features: Features::create_time_series_features(vec![1.0, 2.0, 3.0, 4.0, 5.0]),
|
|
expected_output: ModelPrediction::new(
|
|
"mamba_test".to_string(),
|
|
6.0,
|
|
0.95,
|
|
),
|
|
tolerance: NumericalTolerance::default(),
|
|
description: "MAMBA-2 sequence modeling accuracy".to_string(),
|
|
},
|
|
|
|
// Rainbow DQN Tests
|
|
NumericalTestCase {
|
|
name: "dqn_action_values".to_string(),
|
|
model_type: ModelType::DQN,
|
|
input_features: Features::create_market_state_features(
|
|
vec![100.0, 101.0, 99.5, 100.5], // OHLC
|
|
vec![1000.0, 1500.0], // Volume, Spread
|
|
),
|
|
expected_output: ModelPrediction::Direction(crate::types::DirectionPrediction {
|
|
direction: crate::types::TradeDirection::Buy,
|
|
confidence: 0.87,
|
|
expected_return: 0.02,
|
|
}),
|
|
tolerance: NumericalTolerance::default(),
|
|
description: "Rainbow DQN action value estimation".to_string(),
|
|
},
|
|
|
|
// TLOB Transformer Tests
|
|
NumericalTestCase {
|
|
name: "tlob_order_book_prediction".to_string(),
|
|
model_type: ModelType::TLOB,
|
|
input_features: Features::create_order_book_features(
|
|
vec![(100.0, 1000.0), (100.1, 2000.0)], // Bids
|
|
vec![(100.2, 1500.0), (100.3, 1000.0)], // Asks
|
|
),
|
|
expected_output: ModelPrediction::new(
|
|
"tlob_test".to_string(),
|
|
100.15,
|
|
0.92,
|
|
),
|
|
tolerance: NumericalTolerance {
|
|
absolute: 1e-8,
|
|
relative: 1e-6,
|
|
decision_threshold: 1e-5, // Tighter for price predictions
|
|
},
|
|
description: "TLOB Transformer order book prediction".to_string(),
|
|
},
|
|
|
|
// TFT Tests
|
|
NumericalTestCase {
|
|
name: "tft_temporal_fusion".to_string(),
|
|
model_type: ModelType::TFT,
|
|
input_features: Features::create_multivariate_features(
|
|
vec![
|
|
vec![1.0, 2.0, 3.0],
|
|
vec![4.0, 5.0, 6.0],
|
|
vec![7.0, 8.0, 9.0],
|
|
]
|
|
),
|
|
expected_output: ModelPrediction::new(
|
|
"tft_test".to_string(),
|
|
10.5,
|
|
0.89,
|
|
),
|
|
tolerance: NumericalTolerance::default(),
|
|
description: "TFT temporal fusion accuracy".to_string(),
|
|
},
|
|
]
|
|
}
|
|
}
|
|
|
|
/// Bridge to Python models for reference testing
|
|
pub struct PythonModelBridge {
|
|
// Would connect to Python process running reference implementations
|
|
// For now, this is a placeholder for the interface
|
|
}
|
|
|
|
impl PythonModelBridge {
|
|
pub fn new() -> Result<Self> {
|
|
// Known limitation: Python bridge not implemented
|
|
// Production should use PyO3 for in-process Python interop or
|
|
// spawn subprocess running reference implementation for validation
|
|
Ok(Self {})
|
|
}
|
|
|
|
pub async fn predict(&self, model_type: &ModelType, features: &Features) -> Result<ModelPrediction> {
|
|
// Known limitation: Python reference implementation not connected
|
|
// Production should:
|
|
// 1. Serialize features to JSON/numpy format
|
|
// 2. Call Python reference model via PyO3 or subprocess
|
|
// 3. Compare predictions for numerical accuracy validation
|
|
// 4. Return reference predictions for cross-validation testing
|
|
let _ = (model_type, features); // Suppress unused warnings
|
|
Ok(ModelPrediction::default())
|
|
}
|
|
}
|
|
|
|
/// Generate comprehensive test report
|
|
pub fn generate_test_report(results: &[NumericalTestResult]) -> String {
|
|
let mut report = String::new();
|
|
|
|
report.push_str("# ML Numerical Equivalence Test Report\n\n");
|
|
|
|
let passed = results.iter().filter(|r| r.passed).count();
|
|
let total = results.len();
|
|
let pass_rate = (passed as f64 / total as f64) * 100.0;
|
|
|
|
report.push_str(&format!("## Summary\n"));
|
|
report.push_str(&format!("- **Tests Passed**: {}/{} ({:.1}%)\n", passed, total, pass_rate));
|
|
report.push_str(&format!("- **Production Ready**: {}\n\n", if pass_rate >= 95.0 { "✅ YES" } else { "❌ NO" }));
|
|
|
|
// Latency analysis
|
|
let avg_latency: f64 = results.iter()
|
|
.map(|r| r.latency_ns as f64)
|
|
.sum::<f64>() / results.len() as f64;
|
|
|
|
report.push_str(&format!("## Performance\n"));
|
|
report.push_str(&format!("- **Average Latency**: {:.1}μs\n", avg_latency / 1000.0));
|
|
report.push_str(&format!("- **HFT Ready**: {}\n\n", if avg_latency < 50_000.0 { "✅ Sub-50μs" } else { "⚠️ Above 50μs" }));
|
|
|
|
// Detailed results
|
|
report.push_str("## Detailed Results\n\n");
|
|
for result in results {
|
|
let status = if result.passed { "✅ PASS" } else { "❌ FAIL" };
|
|
report.push_str(&format!("### {} - {}\n", result.test_name, status));
|
|
report.push_str(&format!("- **Absolute Error**: {:.2e}\n", result.absolute_error));
|
|
report.push_str(&format!("- **Relative Error**: {:.2e}\n", result.relative_error));
|
|
report.push_str(&format!("- **Latency**: {:.1}μs\n", result.latency_ns as f64 / 1000.0));
|
|
|
|
if let Some(ref error) = result.error_message {
|
|
report.push_str(&format!("- **Error**: {}\n", error));
|
|
}
|
|
report.push_str("\n");
|
|
}
|
|
|
|
report
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
|
|
#[tokio::test]
|
|
async fn test_numerical_validator_creation() {
|
|
let validator = NumericalValidator::new();
|
|
assert!(!validator.test_cases.is_empty());
|
|
}
|
|
|
|
#[test]
|
|
fn test_tolerance_defaults() {
|
|
let tolerance = NumericalTolerance::default();
|
|
assert!(tolerance.absolute > 0.0);
|
|
assert!(tolerance.relative > 0.0);
|
|
assert!(tolerance.decision_threshold > 0.0);
|
|
}
|
|
|
|
#[test]
|
|
fn test_report_generation() {
|
|
let results = vec![
|
|
NumericalTestResult {
|
|
test_name: "test1".to_string(),
|
|
passed: true,
|
|
rust_output: ModelPrediction::default(),
|
|
python_reference: ModelPrediction::default(),
|
|
absolute_error: 1e-12,
|
|
relative_error: 1e-10,
|
|
latency_ns: 25_000,
|
|
error_message: None,
|
|
}
|
|
];
|
|
|
|
let report = generate_test_report(&results);
|
|
assert!(report.contains("✅ YES"));
|
|
assert!(report.contains("✅ Sub-50μs"));
|
|
}
|
|
} |