828417b523
Integration test now loads imbalance_bar_threshold and ewma_alpha from config/training/dqn-smoketest.toml. Single source of truth for all config values. Production threshold lowered to 0.5 for maximum bar yield. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1081 lines
38 KiB
Rust
1081 lines
38 KiB
Rust
//! MBP-10 Data Loader Helper
|
||
//!
|
||
//! Simple integration layer between DBN MBP-10 parser and OFI feature extraction.
|
||
//! Provides synchronous loading wrapper and snapshot window selection for real-time
|
||
//! Order Flow Imbalance (OFI) calculation.
|
||
//!
|
||
//! # Purpose
|
||
//!
|
||
//! - Load MBP-10 snapshots from DBN files
|
||
//! - Provide snapshot windows for OFI calculation
|
||
//! - Abstract async DBN parser complexity
|
||
//!
|
||
//! # Example
|
||
//!
|
||
//! ```no_run
|
||
//! use ml::features::mbp10_loader::load_mbp10_snapshots_sync;
|
||
//! use ml::features::ofi_calculator::OFICalculator;
|
||
//! use std::path::Path;
|
||
//!
|
||
//! let snapshots = load_mbp10_snapshots_sync(Path::new("test_data/ES.FUT.mbp10.dbn"))?;
|
||
//! let mut calculator = OFICalculator::new();
|
||
//!
|
||
//! for snapshot in &snapshots {
|
||
//! let features = calculator.calculate(snapshot)?;
|
||
//! // Use features for model input
|
||
//! }
|
||
//! ```
|
||
|
||
use data::providers::databento::{dbn_parser::DbnParser, mbp10::{BidAskPair, Mbp10Snapshot}};
|
||
use std::path::Path;
|
||
|
||
use chrono::{DateTime, Utc};
|
||
|
||
use crate::alternative_bars::ImbalanceBarSampler;
|
||
use crate::ofi_calculator::OFICalculator;
|
||
use crate::trades_loader::load_trades_sync;
|
||
use crate::OHLCVBar;
|
||
use crate::MLError;
|
||
|
||
/// Load MBP-10 snapshots from DBN file (synchronous wrapper)
|
||
///
|
||
/// This is a blocking wrapper around the async `DbnParser::parse_mbp10_file()`.
|
||
/// Use this for simple synchronous contexts like testing or batch processing.
|
||
///
|
||
/// # Arguments
|
||
///
|
||
/// * `file_path` - Path to the DBN file containing MBP-10 data
|
||
///
|
||
/// # Returns
|
||
///
|
||
/// * `Ok(Vec<Mbp10Snapshot>)` - All snapshots from the file
|
||
/// * `Err(MLError)` - If file cannot be read or parsed
|
||
///
|
||
/// # Performance
|
||
///
|
||
/// - Target: <10ms for typical 180-day files
|
||
/// - Memory: ~1KB per snapshot (10 levels × 48 bytes)
|
||
///
|
||
pub fn load_mbp10_snapshots_sync(file_path: &Path) -> Result<Vec<Mbp10Snapshot>, MLError> {
|
||
let parser = DbnParser::new().map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create DBN parser: {}", e))
|
||
})?;
|
||
|
||
// Use tokio runtime for async operation
|
||
let runtime = tokio::runtime::Runtime::new().map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create async runtime: {}", e))
|
||
})?;
|
||
|
||
runtime.block_on(async {
|
||
parser
|
||
.parse_mbp10_file(file_path)
|
||
.await
|
||
.map_err(|e| MLError::InsufficientData(format!("Failed to parse MBP-10 file: {}", e)))
|
||
})
|
||
}
|
||
|
||
/// Compute OFI features from an MBP-10 file in a single streaming pass.
|
||
///
|
||
/// This is ~10x faster than `load_mbp10_snapshots_sync` + separate OFI iteration
|
||
/// because it avoids materializing millions of intermediate `Mbp10Snapshot` clones.
|
||
/// The parser feeds snapshots by reference directly to the OFI calculator.
|
||
///
|
||
/// # Arguments
|
||
///
|
||
/// * `file_path` - Path to the .dbn or .dbn.zst MBP-10 file
|
||
///
|
||
/// # Returns
|
||
///
|
||
/// * `Ok(Vec<[f64; 8]>)` - OFI feature arrays (one per aggregated snapshot)
|
||
/// * `Err(MLError)` - If file cannot be read or parsed
|
||
///
|
||
pub fn compute_ofi_from_file(file_path: &Path) -> Result<Vec<[f64; 8]>, MLError> {
|
||
let parser = DbnParser::new().map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create DBN parser: {}", e))
|
||
})?;
|
||
|
||
let mut calculator = OFICalculator::new();
|
||
let mut features = Vec::new();
|
||
|
||
let snapshot_count = parser
|
||
.parse_mbp10_streaming(file_path, 100, |snapshot| {
|
||
if let Ok(f) = calculator.calculate(snapshot) {
|
||
features.push(f.to_array());
|
||
} // Skip failed calculations (e.g. first snapshot with no prev)
|
||
})
|
||
.map_err(|e| MLError::InsufficientData(format!("MBP-10 streaming parse failed: {}", e)))?;
|
||
|
||
tracing::debug!(
|
||
"Computed {} OFI features from {} snapshots in {:?}",
|
||
features.len(),
|
||
snapshot_count,
|
||
file_path.file_name().unwrap_or_default()
|
||
);
|
||
|
||
Ok(features)
|
||
}
|
||
|
||
/// Compute OFI features from MBP-10 file enriched with trade data.
|
||
///
|
||
/// Interleaves trade events (from a `.dbn.zst` trades file) into the MBP-10
|
||
/// streaming pipeline. Before each MBP-10 snapshot is processed, all trades
|
||
/// with timestamps up to that snapshot are fed into the OFI calculator,
|
||
/// populating VPIN, Kyle's Lambda, and trade imbalance with real trade data
|
||
/// instead of leaving them at zero.
|
||
///
|
||
/// Falls back to `compute_ofi_from_file()` if the trade file doesn't exist
|
||
/// or fails to load.
|
||
pub fn compute_ofi_with_trades(
|
||
mbp10_file: &Path,
|
||
trades_file: &Path,
|
||
) -> Result<Vec<[f64; 8]>, MLError> {
|
||
// Load all trades upfront (sorted by timestamp). Trade files are much
|
||
// smaller than MBP-10 (~120 MB vs ~5+ GB), so this is fine in memory.
|
||
let trades = match load_trades_sync(trades_file) {
|
||
Ok(t) => {
|
||
tracing::info!(
|
||
"Loaded {} trades from {} for OFI enrichment",
|
||
t.len(),
|
||
trades_file.file_name().unwrap_or_default().to_string_lossy()
|
||
);
|
||
t
|
||
}
|
||
Err(e) => {
|
||
tracing::warn!(
|
||
"Failed to load trades from {}: {} \u{2014} falling back to MBP-10 only",
|
||
trades_file.display(),
|
||
e
|
||
);
|
||
return compute_ofi_from_file(mbp10_file);
|
||
}
|
||
};
|
||
|
||
let parser = DbnParser::new().map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create DBN parser: {}", e))
|
||
})?;
|
||
|
||
let mut calculator = OFICalculator::new();
|
||
let mut features = Vec::new();
|
||
let mut trade_cursor: usize = 0;
|
||
|
||
let snapshot_count = parser
|
||
.parse_mbp10_streaming(mbp10_file, 100, |snapshot| {
|
||
// Feed all trades with timestamp <= this snapshot's timestamp.
|
||
// This ensures VPIN/Kyle's Lambda/trade_imbalance are populated
|
||
// before the OFI calculator processes the snapshot.
|
||
while trade_cursor < trades.len()
|
||
&& trades[trade_cursor].timestamp <= snapshot.timestamp
|
||
{
|
||
let t = &trades[trade_cursor];
|
||
calculator.feed_trade(t.price, t.volume, t.is_buy);
|
||
trade_cursor += 1;
|
||
}
|
||
|
||
if let Ok(f) = calculator.calculate(snapshot) {
|
||
features.push(f.to_array());
|
||
} // Skip failed calculations (e.g. first snapshot)
|
||
})
|
||
.map_err(|e| MLError::InsufficientData(format!("MBP-10 streaming parse failed: {}", e)))?;
|
||
|
||
tracing::info!(
|
||
"Computed {} OFI features from {} snapshots + {} trades (cursor at {}/{})",
|
||
features.len(),
|
||
snapshot_count,
|
||
trades.len(),
|
||
trade_cursor,
|
||
trades.len()
|
||
);
|
||
|
||
Ok(features)
|
||
}
|
||
|
||
/// Find matching trade file for an MBP-10 file in the trades directory.
|
||
///
|
||
/// Matches by filename stem: `ES.FUT_2024-Q1.dbn.zst` in MBP-10 dir
|
||
/// → looks for `ES.FUT_2024-Q1.dbn.zst` in trades dir (maintaining
|
||
/// the same subdirectory structure).
|
||
fn find_matching_trade_file(mbp10_file: &Path, trades_dir: &Path) -> Option<std::path::PathBuf> {
|
||
let file_name = mbp10_file.file_name()?;
|
||
|
||
// Try exact match: trades_dir/SYMBOL/filename
|
||
// MBP-10 files are typically under mbp10_dir/ES.FUT/ES.FUT_2024-Q1.dbn.zst
|
||
// Trades would be under trades_dir/ES.FUT/ES.FUT_2024-Q1.dbn.zst
|
||
if let Some(parent) = mbp10_file.parent() {
|
||
if let Some(symbol_dir) = parent.file_name() {
|
||
let candidate = trades_dir.join(symbol_dir).join(file_name);
|
||
if candidate.exists() {
|
||
return Some(candidate);
|
||
}
|
||
}
|
||
}
|
||
|
||
// Fallback: flat structure — trades_dir/filename
|
||
let candidate = trades_dir.join(file_name);
|
||
if candidate.exists() {
|
||
return Some(candidate);
|
||
}
|
||
|
||
None
|
||
}
|
||
|
||
/// Load OFI features from all MBP-10 files in a directory using parallel streaming.
|
||
///
|
||
/// Files are processed concurrently with rayon. Each file streams through the DBN parser
|
||
/// and computes OFI features inline — no intermediate `Vec<Mbp10Snapshot>` allocation.
|
||
///
|
||
/// # Arguments
|
||
///
|
||
/// * `mbp10_dir` - Directory containing .dbn or .dbn.zst MBP-10 files
|
||
/// * `file_collector` - Function to collect files from the directory (supports both
|
||
/// flat and recursive collection patterns used by DQN/PPO adapters)
|
||
///
|
||
/// # Returns
|
||
///
|
||
/// * `Some(Vec<[f64; 8]>)` - Concatenated OFI features from all files (file order preserved)
|
||
/// * `None` - If directory doesn't exist, has no files, or all files failed
|
||
///
|
||
#[allow(clippy::cognitive_complexity)]
|
||
pub fn load_ofi_features_parallel<F>(
|
||
mbp10_dir: &Path,
|
||
trades_dir: Option<&Path>,
|
||
file_collector: F,
|
||
) -> Option<Vec<[f64; 8]>>
|
||
where
|
||
F: FnOnce(&Path) -> Vec<std::path::PathBuf>,
|
||
{
|
||
use rayon::prelude::*;
|
||
|
||
if !mbp10_dir.exists() {
|
||
tracing::info!(
|
||
"No MBP10 directory at {}, OFI features will be zero-padded",
|
||
mbp10_dir.display()
|
||
);
|
||
return None;
|
||
}
|
||
|
||
let mbp10_files = file_collector(mbp10_dir);
|
||
if mbp10_files.is_empty() {
|
||
tracing::info!("No MBP10 .dbn files found in {}", mbp10_dir.display());
|
||
return None;
|
||
}
|
||
|
||
tracing::info!(
|
||
"Loading OFI features from {} MBP10 files (parallel streaming{})",
|
||
mbp10_files.len(),
|
||
if trades_dir.is_some() { " + trades" } else { "" }
|
||
);
|
||
|
||
let start = std::time::Instant::now();
|
||
|
||
// Process files in parallel — each file gets its own OFI calculator (independent state).
|
||
// If a trades directory is provided, each MBP-10 file looks for a matching trade file
|
||
// to enrich VPIN, Kyle's Lambda, and trade imbalance with real trade data.
|
||
let per_file_results: Vec<_> = mbp10_files
|
||
.par_iter()
|
||
.map(|file| {
|
||
let result = if let Some(t_dir) = trades_dir {
|
||
if let Some(trade_file) = find_matching_trade_file(file, t_dir) {
|
||
compute_ofi_with_trades(file, &trade_file)
|
||
} else {
|
||
tracing::debug!(
|
||
"No matching trade file for {}, using MBP-10 only",
|
||
file.file_name().unwrap_or_default().to_string_lossy()
|
||
);
|
||
compute_ofi_from_file(file)
|
||
}
|
||
} else {
|
||
compute_ofi_from_file(file)
|
||
};
|
||
|
||
match result {
|
||
Ok(features) => {
|
||
tracing::info!(
|
||
" {} \u{2192} {} OFI features",
|
||
file.file_name().unwrap_or_default().to_string_lossy(),
|
||
features.len()
|
||
);
|
||
features
|
||
}
|
||
Err(e) => {
|
||
tracing::warn!("Failed to compute OFI from {}: {}", file.display(), e);
|
||
Vec::new()
|
||
}
|
||
}
|
||
})
|
||
.collect();
|
||
|
||
// Concatenate results in file order
|
||
let total_features: usize = per_file_results.iter().map(|v| v.len()).sum();
|
||
let mut all_ofi = Vec::with_capacity(total_features);
|
||
for features in per_file_results {
|
||
all_ofi.extend_from_slice(&features);
|
||
}
|
||
|
||
if all_ofi.is_empty() {
|
||
return None;
|
||
}
|
||
|
||
let elapsed = start.elapsed();
|
||
tracing::info!(
|
||
"Computed {} OFI feature vectors from {} files in {:.1}s ({:.0} features/sec)",
|
||
all_ofi.len(),
|
||
mbp10_files.len(),
|
||
elapsed.as_secs_f64(),
|
||
all_ofi.len() as f64 / elapsed.as_secs_f64()
|
||
);
|
||
|
||
Some(all_ofi)
|
||
}
|
||
|
||
/// Trade tick extracted from MBP-10 data.
|
||
///
|
||
/// MBP-10 records include order book updates AND trades. When an MBP-10 record
|
||
/// has `action == b'T'` it represents a trade execution. We extract the price,
|
||
/// size, timestamp, and buy/sell classification for feeding into the
|
||
/// `ImbalanceBarSampler`.
|
||
#[derive(Debug, Clone)]
|
||
pub struct Mbp10Trade {
|
||
/// Trade price (f64, converted from fixed-point)
|
||
pub price: f64,
|
||
/// Trade volume in contracts
|
||
pub volume: f64,
|
||
/// Event timestamp
|
||
pub timestamp: DateTime<Utc>,
|
||
/// True if buyer-initiated (side == 'B'), false otherwise
|
||
pub is_buy: bool,
|
||
}
|
||
|
||
/// Extract trade ticks from MBP-10 snapshots using consecutive snapshot diffs.
|
||
///
|
||
/// MBP-10 snapshots don't directly carry trade action/side info (those live on
|
||
/// the raw `dbn::Mbp10Msg` records). This function uses a heuristic:
|
||
/// when `trade_count` increases between consecutive snapshots, a trade occurred.
|
||
/// Direction is inferred by comparing mid-prices (tick rule).
|
||
///
|
||
/// For higher-fidelity trade extraction, use [`extract_trades_from_dbn_file`]
|
||
/// which reads raw MBP-10 records with their action/side fields.
|
||
pub fn extract_trades_from_snapshots(snapshots: &[Mbp10Snapshot]) -> Vec<Mbp10Trade> {
|
||
let mut trades = Vec::new();
|
||
let mut prev_trade_count: u32 = 0;
|
||
let mut prev_mid: f64 = 0.0;
|
||
|
||
for snap in snapshots {
|
||
let mid = snap.mid_price();
|
||
if mid <= 0.0 {
|
||
continue;
|
||
}
|
||
|
||
let new_trades = snap.trade_count.saturating_sub(prev_trade_count);
|
||
if new_trades > 0 && prev_mid > 0.0 {
|
||
// Tick rule: price up => buy, price down => sell, unchanged => previous direction
|
||
let is_buy = mid >= prev_mid;
|
||
// Best ask size as proxy for trade volume (actual volume not in snapshot)
|
||
let volume = if !snap.levels.is_empty() {
|
||
snap.levels[0].bid_sz.max(1) as f64
|
||
} else {
|
||
1.0
|
||
};
|
||
|
||
let ts_secs = (snap.timestamp / 1_000_000_000) as i64;
|
||
let ts_nanos = (snap.timestamp % 1_000_000_000) as u32;
|
||
if let Some(dt) = DateTime::<Utc>::from_timestamp(ts_secs, ts_nanos) {
|
||
trades.push(Mbp10Trade {
|
||
price: mid,
|
||
volume,
|
||
timestamp: dt,
|
||
is_buy,
|
||
});
|
||
}
|
||
}
|
||
|
||
prev_trade_count = snap.trade_count;
|
||
prev_mid = mid;
|
||
}
|
||
|
||
trades
|
||
}
|
||
|
||
/// Extract trade ticks directly from a DBN MBP-10 file.
|
||
///
|
||
/// Reads raw `Mbp10Msg` records and filters for trade actions (`action == b'T'`).
|
||
/// This provides higher-fidelity trades than [`extract_trades_from_snapshots`]
|
||
/// because it uses the actual action and side fields from each record.
|
||
pub fn extract_trades_from_dbn_file(file_path: &Path) -> Result<Vec<Mbp10Trade>, MLError> {
|
||
use dbn::decode::{DbnDecoder, DecodeRecordRef};
|
||
use dbn::RecordRefEnum;
|
||
use std::fs::File;
|
||
use std::io::BufReader;
|
||
|
||
let file = File::open(file_path).map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to open MBP-10 file {:?}: {}", file_path, e))
|
||
})?;
|
||
|
||
let is_zstd = file_path.to_string_lossy().ends_with(".dbn.zst");
|
||
let reader: Box<dyn std::io::Read> = if is_zstd {
|
||
Box::new(
|
||
zstd::Decoder::new(BufReader::new(file)).map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create zstd decoder: {}", e))
|
||
})?,
|
||
)
|
||
} else {
|
||
Box::new(BufReader::new(file))
|
||
};
|
||
|
||
let mut decoder = DbnDecoder::new(reader).map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to create DBN decoder: {}", e))
|
||
})?;
|
||
|
||
let mut trades = Vec::new();
|
||
let mut last_mid: f64 = 0.0;
|
||
|
||
loop {
|
||
match decoder.decode_record_ref() {
|
||
Ok(Some(record)) => {
|
||
let record_enum = record.as_enum().map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to convert record: {}", e))
|
||
})?;
|
||
|
||
if let RecordRefEnum::Mbp10(mbp10) = record_enum {
|
||
let action = mbp10.action as u8;
|
||
|
||
// Compute mid-price from this record's top-of-book levels
|
||
// MBP-10 records carry 10 levels; index 0 is best bid/ask
|
||
let bid_px = BidAskPair::price_to_f64(mbp10.levels[0].bid_px);
|
||
let ask_px = BidAskPair::price_to_f64(mbp10.levels[0].ask_px);
|
||
let mid = if bid_px > 0.0 && ask_px > 0.0 {
|
||
(bid_px + ask_px) / 2.0
|
||
} else {
|
||
last_mid
|
||
};
|
||
if mid > 0.0 {
|
||
last_mid = mid;
|
||
}
|
||
|
||
// Trade action: 'T' (0x54) = trade execution
|
||
if action == b'T' {
|
||
let price = mbp10.price as f64 * 1e-9;
|
||
let volume = mbp10.size as f64;
|
||
let is_buy = mbp10.side == b'B' as i8;
|
||
|
||
let ts_secs = (mbp10.hd.ts_event / 1_000_000_000) as i64;
|
||
let ts_nanos = (mbp10.hd.ts_event % 1_000_000_000) as u32;
|
||
if let Some(dt) = DateTime::<Utc>::from_timestamp(ts_secs, ts_nanos) {
|
||
if price > 0.0 && volume > 0.0 {
|
||
trades.push(Mbp10Trade {
|
||
price,
|
||
volume,
|
||
timestamp: dt,
|
||
is_buy,
|
||
});
|
||
}
|
||
}
|
||
}
|
||
}
|
||
}
|
||
Ok(None) => break,
|
||
Err(e) => {
|
||
return Err(MLError::InsufficientData(format!(
|
||
"DBN decode error: {}",
|
||
e
|
||
)));
|
||
}
|
||
}
|
||
}
|
||
|
||
Ok(trades)
|
||
}
|
||
|
||
/// Recursively collect all .dbn and .dbn.zst files from a directory.
|
||
fn collect_dbn_files(dir: &Path) -> Vec<std::path::PathBuf> {
|
||
let mut files = Vec::new();
|
||
if let Ok(entries) = std::fs::read_dir(dir) {
|
||
for entry in entries.flatten() {
|
||
let path = entry.path();
|
||
if path.is_dir() {
|
||
files.extend(collect_dbn_files(&path));
|
||
} else if path.extension().and_then(|s| s.to_str()) == Some("dbn")
|
||
|| path.to_string_lossy().ends_with(".dbn.zst")
|
||
{
|
||
files.push(path);
|
||
}
|
||
}
|
||
}
|
||
files
|
||
}
|
||
|
||
// ══════════════════════════════════════════════════════════════════════════════
|
||
// Imbalance Bar Cache
|
||
//
|
||
// Computing imbalance bars from MBP-10 data requires decompressing ~19GB of
|
||
// .dbn.zst files (~450s). This cache stores the resulting `Vec<OHLCVBar>` to
|
||
// `/tmp/.foxhunt_imbalance_cache/<hash>.bars.bin` using bincode.
|
||
//
|
||
// Cache key: DefaultHasher of (mbp10_dir, symbol, threshold bits, alpha bits).
|
||
// Invalidation: cache mtime vs. newest .dbn/.dbn.zst file in the source dir.
|
||
// ══════════════════════════════════════════════════════════════════════════════
|
||
|
||
/// Compute a cache path for an imbalance bar set.
|
||
fn imbalance_cache_path(mbp10_dir: &Path, symbol: &str, threshold: f64, alpha: f64) -> std::path::PathBuf {
|
||
use std::collections::hash_map::DefaultHasher;
|
||
use std::hash::{Hash, Hasher};
|
||
let mut hasher = DefaultHasher::new();
|
||
mbp10_dir.hash(&mut hasher);
|
||
symbol.hash(&mut hasher);
|
||
threshold.to_bits().hash(&mut hasher);
|
||
alpha.to_bits().hash(&mut hasher);
|
||
let hash = hasher.finish();
|
||
std::path::PathBuf::from(format!("/tmp/.foxhunt_imbalance_cache/{:016x}.bars.bin", hash))
|
||
}
|
||
|
||
/// Check whether the cache file is newer than every `.dbn`/`.dbn.zst` file under `source_dir`.
|
||
fn is_imbalance_cache_valid(cache: &Path, source_dir: &Path) -> bool {
|
||
let cache_mtime = match std::fs::metadata(cache) {
|
||
Ok(m) => match m.modified() {
|
||
Ok(t) => t,
|
||
Err(_) => return false,
|
||
},
|
||
Err(_) => return false,
|
||
};
|
||
|
||
fn walk_dbn_mtimes(dir: &Path, newest: &mut std::time::SystemTime) -> std::io::Result<()> {
|
||
for entry in std::fs::read_dir(dir)? {
|
||
let entry = entry?;
|
||
let path = entry.path();
|
||
if path.is_dir() {
|
||
walk_dbn_mtimes(&path, newest)?;
|
||
} else if path.extension().and_then(|s| s.to_str()) == Some("dbn")
|
||
|| path.to_string_lossy().ends_with(".dbn.zst")
|
||
{
|
||
if let Ok(meta) = std::fs::metadata(&path) {
|
||
if let Ok(mtime) = meta.modified() {
|
||
if mtime > *newest {
|
||
*newest = mtime;
|
||
}
|
||
}
|
||
}
|
||
}
|
||
}
|
||
Ok(())
|
||
}
|
||
|
||
let mut newest_source = std::time::SystemTime::UNIX_EPOCH;
|
||
if walk_dbn_mtimes(source_dir, &mut newest_source).is_err() {
|
||
return false;
|
||
}
|
||
|
||
// Cache is valid if it was created after all source files were last modified
|
||
cache_mtime >= newest_source
|
||
}
|
||
|
||
/// Load MBP-10 files and convert to imbalance bars via `ImbalanceBarSampler`.
|
||
///
|
||
/// Scans `mbp10_dir/symbol/` for `.dbn` and `.dbn.zst` files, loads each,
|
||
/// extracts trades (action == 'T'), feeds into adaptive `ImbalanceBarSampler`.
|
||
///
|
||
/// Results are cached to `/tmp/.foxhunt_imbalance_cache/<hash>.bars.bin`.
|
||
/// Subsequent calls with the same parameters return in <1s if no source files
|
||
/// have changed.
|
||
///
|
||
/// Returns sorted `OHLCVBar`s or error if directory doesn't exist.
|
||
/// No fallback -- fails loudly if the data isn't there.
|
||
pub fn mbp10_to_imbalance_bars(
|
||
mbp10_dir: &Path,
|
||
symbol: &str,
|
||
threshold: f64,
|
||
ewma_alpha: f64,
|
||
) -> Result<Vec<OHLCVBar>, MLError> {
|
||
let symbol_dir = mbp10_dir.join(symbol);
|
||
if !symbol_dir.exists() {
|
||
return Err(MLError::InsufficientData(format!(
|
||
"MBP-10 directory not found: {}. Set data_source='ohlcv' or provide MBP-10 data.",
|
||
symbol_dir.display()
|
||
)));
|
||
}
|
||
|
||
// ── Cache check ──────────────────────────────────────────────────────
|
||
let cache_file = imbalance_cache_path(mbp10_dir, symbol, threshold, ewma_alpha);
|
||
if is_imbalance_cache_valid(&cache_file, &symbol_dir) {
|
||
let load_start = std::time::Instant::now();
|
||
match load_imbalance_cache(&cache_file) {
|
||
Ok(bars) => {
|
||
let elapsed_ms = load_start.elapsed().as_secs_f64() * 1000.0;
|
||
tracing::info!(
|
||
"Imbalance bar cache HIT: loaded {} bars in {:.1}ms",
|
||
bars.len(),
|
||
elapsed_ms
|
||
);
|
||
return Ok(bars);
|
||
}
|
||
Err(e) => {
|
||
tracing::warn!("Imbalance bar cache read failed (recomputing): {}", e);
|
||
}
|
||
}
|
||
}
|
||
|
||
// ── Cache miss — compute from scratch ────────────────────────────────
|
||
let mut dbn_files = collect_dbn_files(&symbol_dir);
|
||
dbn_files.sort();
|
||
|
||
if dbn_files.is_empty() {
|
||
return Err(MLError::InsufficientData(format!(
|
||
"No .dbn/.dbn.zst files found in {}",
|
||
symbol_dir.display()
|
||
)));
|
||
}
|
||
|
||
tracing::info!(
|
||
"Imbalance bar cache MISS: computing from {} MBP-10 files...",
|
||
dbn_files.len()
|
||
);
|
||
|
||
// Collect all trades from all files
|
||
let mut all_trades: Vec<Mbp10Trade> = Vec::new();
|
||
for file_path in &dbn_files {
|
||
match extract_trades_from_dbn_file(file_path) {
|
||
Ok(trades) => {
|
||
tracing::debug!(
|
||
" {} -> {} trades",
|
||
file_path.file_name().unwrap_or_default().to_string_lossy(),
|
||
trades.len()
|
||
);
|
||
all_trades.extend(trades);
|
||
}
|
||
Err(e) => {
|
||
tracing::warn!(
|
||
"Failed to extract trades from {}: {}",
|
||
file_path.display(),
|
||
e
|
||
);
|
||
}
|
||
}
|
||
}
|
||
|
||
if all_trades.is_empty() {
|
||
return Err(MLError::InsufficientData(format!(
|
||
"No trade ticks extracted from MBP-10 files in {}. \
|
||
Files may contain only book updates without trade events.",
|
||
symbol_dir.display()
|
||
)));
|
||
}
|
||
|
||
// Sort by timestamp (files should be chronological but ensure correctness)
|
||
all_trades.sort_by(|a, b| a.timestamp.cmp(&b.timestamp));
|
||
|
||
tracing::info!(
|
||
"Extracted {} trade ticks, feeding into ImbalanceBarSampler (threshold={}, alpha={})",
|
||
all_trades.len(),
|
||
threshold,
|
||
ewma_alpha,
|
||
);
|
||
|
||
// Initialize sampler with first trade
|
||
let first = &all_trades[0];
|
||
let mut sampler =
|
||
ImbalanceBarSampler::new_with_ewma(first.price, threshold, first.timestamp, ewma_alpha);
|
||
|
||
let mut bars: Vec<OHLCVBar> = Vec::new();
|
||
for trade in &all_trades {
|
||
if let Some(bar) = sampler.update(trade.price, trade.volume, trade.timestamp) {
|
||
bars.push(bar);
|
||
}
|
||
}
|
||
|
||
// Sort by timestamp (should already be ordered)
|
||
bars.sort_by(|a, b| a.timestamp.cmp(&b.timestamp));
|
||
|
||
tracing::info!(
|
||
"ImbalanceBarSampler produced {} bars from {} trade ticks (ratio 1:{:.0})",
|
||
bars.len(),
|
||
all_trades.len(),
|
||
if bars.is_empty() {
|
||
0.0
|
||
} else {
|
||
all_trades.len() as f64 / bars.len() as f64
|
||
},
|
||
);
|
||
|
||
// ── Save to cache ────────────────────────────────────────────────────
|
||
if let Err(e) = save_imbalance_cache(&cache_file, &bars) {
|
||
tracing::warn!("Failed to write imbalance bar cache (non-fatal): {}", e);
|
||
} else {
|
||
tracing::info!("Cached {} bars to {}", bars.len(), cache_file.display());
|
||
}
|
||
|
||
Ok(bars)
|
||
}
|
||
|
||
/// Deserialize cached imbalance bars from a bincode file.
|
||
fn load_imbalance_cache(path: &Path) -> Result<Vec<OHLCVBar>, MLError> {
|
||
use std::io::BufReader;
|
||
let file = std::fs::File::open(path).map_err(|e| {
|
||
MLError::InsufficientData(format!("Failed to open imbalance cache {:?}: {}", path, e))
|
||
})?;
|
||
let reader = BufReader::new(file);
|
||
bincode::deserialize_from(reader).map_err(|e| {
|
||
MLError::InsufficientData(format!(
|
||
"Failed to deserialize imbalance cache {:?}: {}",
|
||
path, e
|
||
))
|
||
})
|
||
}
|
||
|
||
/// Serialize imbalance bars to a bincode file, creating the cache directory if needed.
|
||
fn save_imbalance_cache(path: &Path, bars: &[OHLCVBar]) -> Result<(), MLError> {
|
||
use std::io::BufWriter;
|
||
if let Some(parent) = path.parent() {
|
||
std::fs::create_dir_all(parent).map_err(|e| {
|
||
MLError::InsufficientData(format!(
|
||
"Failed to create imbalance cache dir {:?}: {}",
|
||
parent, e
|
||
))
|
||
})?;
|
||
}
|
||
let file = std::fs::File::create(path).map_err(|e| {
|
||
MLError::InsufficientData(format!(
|
||
"Failed to create imbalance cache {:?}: {}",
|
||
path, e
|
||
))
|
||
})?;
|
||
let writer = BufWriter::new(file);
|
||
bincode::serialize_into(writer, bars).map_err(|e| {
|
||
MLError::InsufficientData(format!(
|
||
"Failed to serialize imbalance cache {:?}: {}",
|
||
path, e
|
||
))
|
||
})
|
||
}
|
||
|
||
/// Get snapshots for a specific timestamp window
|
||
///
|
||
/// Returns a slice of snapshots starting from the first snapshot at or before
|
||
/// the target timestamp. Used for calculating OFI features that require
|
||
/// forward-looking context (e.g., next N snapshots after a given timestamp).
|
||
///
|
||
/// # Arguments
|
||
///
|
||
/// * `snapshots` - All available snapshots (must be sorted by timestamp)
|
||
/// * `target_ts` - Target timestamp (nanoseconds since Unix epoch)
|
||
/// * `window_size` - Number of snapshots to include in window
|
||
///
|
||
/// # Returns
|
||
///
|
||
/// Slice of up to `window_size` snapshots starting from the first snapshot
|
||
/// with timestamp >= `target_ts`. Returns empty slice if no snapshots exist
|
||
/// at or after the target timestamp.
|
||
///
|
||
/// # Algorithm
|
||
///
|
||
/// - Simple linear search (TODO: optimize with binary search for large datasets)
|
||
/// - Returns snapshots starting from the first one with timestamp >= `target_ts`
|
||
///
|
||
/// # Example
|
||
///
|
||
/// ```ignore
|
||
/// // Snapshots at times: [1000, 2000, 3000, 4000]
|
||
/// // target_ts = 2000, window_size = 2
|
||
/// // Returns: [2000, 3000]
|
||
/// ```
|
||
///
|
||
pub fn get_snapshots_for_timestamp(
|
||
snapshots: &[Mbp10Snapshot],
|
||
target_ts: u64,
|
||
window_size: usize,
|
||
) -> &[Mbp10Snapshot] {
|
||
if snapshots.is_empty() {
|
||
return &[];
|
||
}
|
||
|
||
// Binary search for the first snapshot with timestamp >= target_ts
|
||
// Snapshots are pre-sorted by timestamp (from DBN file ordering)
|
||
let idx = snapshots.partition_point(|s| s.timestamp < target_ts);
|
||
if idx >= snapshots.len() {
|
||
return &[]; // Target timestamp is after all snapshots
|
||
}
|
||
let end_idx = (idx + window_size).min(snapshots.len());
|
||
&snapshots[idx..end_idx]
|
||
}
|
||
|
||
/// Get the most recent N snapshots ending at the given index
|
||
///
|
||
/// Useful for calculating features that require a rolling window of recent data.
|
||
///
|
||
/// # Arguments
|
||
///
|
||
/// * `snapshots` - All available snapshots
|
||
/// * `end_idx` - Index of the last snapshot to include (exclusive)
|
||
/// * `window_size` - Number of snapshots to include
|
||
///
|
||
/// # Returns
|
||
///
|
||
/// Slice of up to `window_size` snapshots ending at `end_idx`
|
||
///
|
||
pub fn get_recent_snapshots(
|
||
snapshots: &[Mbp10Snapshot],
|
||
end_idx: usize,
|
||
window_size: usize,
|
||
) -> &[Mbp10Snapshot] {
|
||
if snapshots.is_empty() || end_idx == 0 {
|
||
return &[];
|
||
}
|
||
|
||
let actual_end = end_idx.min(snapshots.len());
|
||
let start_idx = actual_end.saturating_sub(window_size);
|
||
|
||
&snapshots[start_idx..actual_end]
|
||
}
|
||
|
||
#[cfg(test)]
|
||
mod tests {
|
||
use super::*;
|
||
use data::providers::databento::mbp10::BidAskPair;
|
||
|
||
fn create_test_snapshot(timestamp: u64, bid_px: i64) -> Mbp10Snapshot {
|
||
let levels = vec![BidAskPair {
|
||
bid_px,
|
||
bid_sz: 100,
|
||
bid_ct: 5,
|
||
ask_px: bid_px + 1000000000, // 1 tick higher
|
||
ask_sz: 120,
|
||
ask_ct: 6,
|
||
}];
|
||
|
||
Mbp10Snapshot::new("ES.FUT".to_owned(), timestamp, levels, 0, 0)
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_for_timestamp_exact_match() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
create_test_snapshot(3000, 150020000000000),
|
||
create_test_snapshot(4000, 150030000000000),
|
||
];
|
||
|
||
let window = get_snapshots_for_timestamp(&snapshots, 2000, 2);
|
||
assert_eq!(window.len(), 2);
|
||
assert_eq!(window[0].timestamp, 2000);
|
||
assert_eq!(window[1].timestamp, 3000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_for_timestamp_between() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
create_test_snapshot(3000, 150020000000000),
|
||
];
|
||
|
||
// Target timestamp between 2000 and 3000
|
||
let window = get_snapshots_for_timestamp(&snapshots, 2500, 2);
|
||
assert_eq!(window.len(), 1); // Only snapshot at 3000 remains
|
||
assert_eq!(window[0].timestamp, 3000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_for_timestamp_empty() {
|
||
let snapshots: Vec<Mbp10Snapshot> = vec![];
|
||
let window = get_snapshots_for_timestamp(&snapshots, 2000, 2);
|
||
assert_eq!(window.len(), 0);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_for_timestamp_before_all() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(2000, 150000000000000),
|
||
create_test_snapshot(3000, 150010000000000),
|
||
];
|
||
|
||
// Target is before all snapshots, should return first 2 snapshots
|
||
let window = get_snapshots_for_timestamp(&snapshots, 1000, 2);
|
||
assert_eq!(window.len(), 2);
|
||
assert_eq!(window[0].timestamp, 2000);
|
||
assert_eq!(window[1].timestamp, 3000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_for_timestamp_after_all() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
];
|
||
|
||
let window = get_snapshots_for_timestamp(&snapshots, 5000, 2);
|
||
assert_eq!(window.len(), 0); // No snapshots after target
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_snapshots_window_clipping() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
create_test_snapshot(3000, 150020000000000),
|
||
];
|
||
|
||
// Request 10 snapshots but only 2 available at/after target (2000, 3000)
|
||
let window = get_snapshots_for_timestamp(&snapshots, 2000, 10);
|
||
assert_eq!(window.len(), 2); // 2 snapshots at/after 2000: [2000, 3000]
|
||
assert_eq!(window[0].timestamp, 2000);
|
||
assert_eq!(window[1].timestamp, 3000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_recent_snapshots_full_window() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
create_test_snapshot(3000, 150020000000000),
|
||
create_test_snapshot(4000, 150030000000000),
|
||
];
|
||
|
||
let window = get_recent_snapshots(&snapshots, 3, 2);
|
||
assert_eq!(window.len(), 2);
|
||
assert_eq!(window[0].timestamp, 2000);
|
||
assert_eq!(window[1].timestamp, 3000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_recent_snapshots_partial_window() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
];
|
||
|
||
// Request 5 snapshots but only 2 available
|
||
let window = get_recent_snapshots(&snapshots, 2, 5);
|
||
assert_eq!(window.len(), 2);
|
||
assert_eq!(window[0].timestamp, 1000);
|
||
assert_eq!(window[1].timestamp, 2000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_recent_snapshots_empty() {
|
||
let snapshots: Vec<Mbp10Snapshot> = vec![];
|
||
let window = get_recent_snapshots(&snapshots, 0, 2);
|
||
assert_eq!(window.len(), 0);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_recent_snapshots_zero_end_idx() {
|
||
let snapshots = vec![
|
||
create_test_snapshot(1000, 150000000000000),
|
||
create_test_snapshot(2000, 150010000000000),
|
||
];
|
||
|
||
let window = get_recent_snapshots(&snapshots, 0, 2);
|
||
assert_eq!(window.len(), 0);
|
||
}
|
||
|
||
#[test]
|
||
fn test_get_recent_snapshots_one_element() {
|
||
let snapshots = vec![create_test_snapshot(1000, 150000000000000)];
|
||
|
||
let window = get_recent_snapshots(&snapshots, 1, 1);
|
||
assert_eq!(window.len(), 1);
|
||
assert_eq!(window[0].timestamp, 1000);
|
||
}
|
||
|
||
#[test]
|
||
fn test_extract_trades_from_snapshots() {
|
||
// Create snapshots with increasing trade_count and rising prices
|
||
let snapshots = vec![
|
||
{
|
||
let mut s = create_test_snapshot(1_000_000_000, 150000000000000); // 150.0
|
||
s.trade_count = 0;
|
||
s
|
||
},
|
||
{
|
||
let mut s = create_test_snapshot(2_000_000_000, 150010000000000); // 150.01
|
||
s.trade_count = 1;
|
||
s
|
||
},
|
||
{
|
||
let mut s = create_test_snapshot(3_000_000_000, 150020000000000); // 150.02
|
||
s.trade_count = 3;
|
||
s
|
||
},
|
||
{
|
||
let mut s = create_test_snapshot(4_000_000_000, 149990000000000); // 149.99
|
||
s.trade_count = 4;
|
||
s
|
||
},
|
||
];
|
||
|
||
let trades = extract_trades_from_snapshots(&snapshots);
|
||
// 3 transitions where trade_count increases (snap 0->1, 1->2, 2->3)
|
||
assert_eq!(trades.len(), 3);
|
||
// First trade: price went up => is_buy=true
|
||
assert!(trades[0].is_buy);
|
||
// Third trade: price went down => is_buy=false
|
||
assert!(!trades[2].is_buy);
|
||
}
|
||
|
||
#[test]
|
||
fn test_mbp10_to_imbalance_bars_missing_dir() {
|
||
let result = mbp10_to_imbalance_bars(
|
||
Path::new("/nonexistent/path"),
|
||
"ES.FUT",
|
||
100.0,
|
||
0.1,
|
||
);
|
||
assert!(result.is_err());
|
||
let err_msg = format!("{}", result.unwrap_err());
|
||
assert!(err_msg.contains("MBP-10 directory not found"), "Error: {err_msg}");
|
||
}
|
||
|
||
/// Integration test: loads real MBP-10 data and produces imbalance bars.
|
||
/// Reads threshold/alpha from config/training/dqn-smoketest.toml (single source of truth).
|
||
/// Requires test_data/futures-baseline-mbp10/ES.FUT/ with .dbn.zst files.
|
||
/// Run with: cargo test -p ml-features --lib -- test_mbp10_to_imbalance_bars_real --ignored --nocapture
|
||
#[test]
|
||
#[ignore]
|
||
fn test_mbp10_to_imbalance_bars_real() {
|
||
// Resolve workspace root from CARGO_MANIFEST_DIR
|
||
let manifest = std::env::var("CARGO_MANIFEST_DIR").unwrap();
|
||
let workspace = std::path::Path::new(&manifest)
|
||
.parent() // crates/
|
||
.and_then(|p| p.parent()) // repo root
|
||
.expect("Failed to find workspace root");
|
||
|
||
// Load threshold/alpha from TOML config (no hardcoded values)
|
||
let toml_path = workspace.join("config/training/dqn-smoketest.toml");
|
||
let toml_str = std::fs::read_to_string(&toml_path)
|
||
.unwrap_or_else(|_| panic!("Failed to read {}", toml_path.display()));
|
||
let toml: toml::Value = toml_str.parse()
|
||
.unwrap_or_else(|e| panic!("Failed to parse TOML: {e}"));
|
||
let training = toml.get("training").expect("[training] section missing");
|
||
let threshold = training.get("imbalance_bar_threshold")
|
||
.and_then(|v| v.as_float()).unwrap_or(1.0);
|
||
let alpha = training.get("imbalance_bar_ewma_alpha")
|
||
.and_then(|v| v.as_float()).unwrap_or(0.1);
|
||
println!("Config: threshold={threshold}, alpha={alpha} (from dqn-smoketest.toml)");
|
||
|
||
let mbp10_dir = workspace.join("test_data/futures-baseline-mbp10");
|
||
assert!(mbp10_dir.exists(), "MBP-10 test data not found at {}", mbp10_dir.display());
|
||
|
||
let bars = mbp10_to_imbalance_bars(&mbp10_dir, "ES.FUT", threshold, alpha)
|
||
.expect("mbp10_to_imbalance_bars must succeed with real data");
|
||
|
||
assert!(!bars.is_empty(), "Must produce imbalance bars from MBP-10 data");
|
||
println!("Produced {} imbalance bars from MBP-10 data", bars.len());
|
||
println!(" First bar: {:?}", bars.first().unwrap().timestamp);
|
||
println!(" Last bar: {:?}", bars.last().unwrap().timestamp);
|
||
|
||
// Verify time-ordering
|
||
for w in bars.windows(2) {
|
||
assert!(
|
||
w[0].timestamp <= w[1].timestamp,
|
||
"Bars must be time-ordered: {:?} > {:?}",
|
||
w[0].timestamp, w[1].timestamp
|
||
);
|
||
}
|
||
|
||
// Verify reasonable bar properties
|
||
for bar in &bars {
|
||
assert!(bar.high >= bar.low, "high must >= low");
|
||
assert!(bar.volume > 0.0, "volume must be positive");
|
||
assert!(bar.close > 0.0, "close must be positive");
|
||
}
|
||
println!("All {} bars valid: time-ordered, positive volume, valid OHLC", bars.len());
|
||
}
|
||
}
|