feat: numba #143

Satvik-Singh192 · 2025-11-15T14:15:09Z

Description

This pull request implements performance optimizations for the core engine by introducing accelerated execution paths using Numba and Cython, along with a full benchmarking suite.

Key Enhancements

Profiled the existing backtest pipeline to identify hot paths (core loop & PnL computation).
Added numba_opt.py implementing Numba JIT-accelerated versions of critical functions.
Created benchmarks/bench_opt.py to benchmark:
- Vanilla (pure Python)
- Numba-optimized
- Cython-optimized
Added CI benchmark workflow (manual/canary) that publishes benchmark artifacts for easy inspection.

These improvements significantly enhance performance for large-scale experiments and production-grade scenarios.

Semver Changes

Patch (bug fix, no new features)
Minor (new features, no breaking changes)
Major (breaking changes)

Issues

This pull request closes the following issue(s):

perf: accelerate core engine with Numba/Cython and provide benchmarking #92 — perf: accelerate core engine with Numba/Cython and provide benchmarking

Checklist

I have read the Contributing Guidelines.
I have profiled and validated the performance improvements.
I have added and verified Numba and Cython accelerated implementations.
I have included benchmark scripts and ensured stable reproducibility.
I have added relevant documentation and example usage where necessary.
All tests pass and linting tools (ruff/black) have been run successfully.

netlify · 2025-11-15T14:15:14Z

✅ Deploy Preview for strong-duckanoo-898b2c ready!

Name	Link
🔨 Latest commit	`818058b`
🔍 Latest deploy log	https://app.netlify.com/projects/strong-duckanoo-898b2c/deploys/69188aef7d3bc5000810fe56
😎 Deploy Preview	https://deploy-preview-143--strong-duckanoo-898b2c.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

To edit notification comments on pull requests, go to your Netlify project configuration.

Satvik-Singh192 · 2025-11-15T14:15:26Z

@ayushkrtiwari please review my Pull request sir

Copilot

Pull Request Overview

This PR introduces performance optimizations for the backtest engine by adding Numba JIT-compiled and Cython-compiled alternatives to critical computation functions, along with comprehensive benchmarking and profiling tools.

Key Changes:

Added Numba-accelerated implementations of core backtest operations (strategy returns, turnover, portfolio value calculations)
Introduced Cython-optimized versions for additional performance comparison
Created benchmarking suite to measure performance gains across different dataset sizes
Added profiling script to identify performance bottlenecks

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 13 comments.

Show a summary per file

File	Description
`src/quant_research_starter/backtest/numba_opt.py`	Implements JIT-compiled versions of backtest operations with Numba decorators and fallback support
`src/quant_research_starter/backtest/cython_opt.pyx`	Provides Cython-optimized implementations for strategy returns and turnover calculations
`src/quant_research_starter/benchmarks/bench_opt.py`	Benchmarking script comparing vanilla, Numba, and Cython implementations across multiple dataset sizes
`src/quant_research_starter/backtest/profile_backtest.py`	Profiling utility to identify performance hotspots in the backtest pipeline
`src/quant_research_starter/backtest/setup_cython.py`	Build configuration for compiling Cython extensions
`.github/workflows/benchmark.yml`	CI workflow for automated benchmark execution and artifact storage

Comments suppressed due to low confidence (1)

src/quant_research_starter/benchmarks/bench_opt.py:77

This assignment to 'current_weights' is unnecessary as it is redefined before this value is used.

    current_weights = np.zeros(n_assets, dtype=np.float64)

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2025-11-15T14:26:46Z

src/quant_research_starter/backtest/numba_opt.py

+
+    for i in prange(n_days):
+        ret_sum = 0.0
+        for j in prange(n_assets):


Nested prange usage in Numba may cause performance issues. Using prange for both the outer loop (line 32) and inner loop (line 34) can lead to thread contention and may not provide the expected parallelization benefits. Consider using prange only for the outer loop and regular range for the inner loop, or using a flattened parallelization strategy.

Suggested change

for j in prange(n_assets):

for j in range(n_assets):

Copilot · 2025-11-15T14:26:46Z

src/quant_research_starter/backtest/profile_backtest.py

+    stats.sort_stats("tottime")
+    stats.print_stats(20)
+
+    print("\nTop 20 functions by total time:")
+    s2 = StringIO()
+    stats = pstats.Stats(profiler, stream=s2)
+    stats.sort_stats("tottime")
+    stats.print_stats(20)
+    print(s2.getvalue())


Duplicate profiler statistics generation. Lines 36-41 and lines 47-51 both generate and print statistics sorted by "tottime", but the first set (lines 43-44) regenerates the stats object unnecessarily. The second pstats.Stats creation on line 48 overwrites the previous stats configuration, making lines 43-44 redundant. Consider removing lines 43-44 or restructuring to avoid duplication.

Copilot · 2025-11-15T14:26:47Z

src/quant_research_starter/benchmarks/bench_opt.py

+    return prices, signals
+
+
+def benchmark_vanilla(prices: pd.DataFrame, signals: pd.DataFrame) -> float:


The return type hint for benchmark_vanilla is incorrect. The function returns a tuple (elapsed, results) but is annotated to return float. The return type should be tuple[float, dict] or similar.

Suggested change

def benchmark_vanilla(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

def benchmark_vanilla(prices: pd.DataFrame, signals: pd.DataFrame) -> tuple[float, dict]:

Copilot · 2025-11-15T14:26:47Z

src/quant_research_starter/benchmarks/bench_opt.py

+    return rank_based_weights(signals, max_leverage, long_pct, short_pct)
+
+
+def benchmark_cython(prices: pd.DataFrame, signals: pd.DataFrame) -> float:


The return type hint for benchmark_cython is incorrect. The function returns a tuple (elapsed, results) or (None, None) but is annotated to return float. The return type should be tuple[float | None, dict | None] or similar.

Suggested change

def benchmark_cython(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

def benchmark_cython(prices: pd.DataFrame, signals: pd.DataFrame) -> tuple[float | None, dict | None]:

Copilot · 2025-11-15T14:26:47Z

src/quant_research_starter/benchmarks/bench_opt.py

+    current_weights = np.zeros(n_assets, dtype=np.float64)
+    for date in returns_df.index:
+        signal_row = aligned_signals.loc[date].values.astype(np.float64)
+        weights = compute_rank_weights_numba(signal_row, 1.0, 0.9, 0.1)
+        current_weights = weights.copy()
+        weights_list.append(current_weights)


Inefficient sequential use of .copy(). Line 81 uses .copy() but the result weights is immediately assigned to current_weights, making the copy unnecessary. The code should either use current_weights = weights without the copy, or remove the intermediate weights variable entirely.

Suggested change

current_weights = np.zeros(n_assets, dtype=np.float64)

for date in returns_df.index:

signal_row = aligned_signals.loc[date].values.astype(np.float64)

weights = compute_rank_weights_numba(signal_row, 1.0, 0.9, 0.1)

current_weights = weights.copy()

weights_list.append(current_weights)

# current_weights = np.zeros(n_assets, dtype=np.float64)

for date in returns_df.index:

signal_row = aligned_signals.loc[date].values.astype(np.float64)

weights = compute_rank_weights_numba(signal_row, 1.0, 0.9, 0.1)

weights_list.append(weights)

Copilot · 2025-11-15T14:26:48Z

src/quant_research_starter/backtest/numba_opt.py

+    n_assets = len(signals)
+    weights = np.zeros(n_assets)
+
+    valid_mask = np.zeros(n_assets, dtype=np.bool_)


Deprecated np.bool_ dtype usage. Line 93 uses dtype=np.bool_ which is deprecated in NumPy 1.20+ and will be removed. Use dtype=np.bool or dtype=bool instead.

Suggested change

valid_mask = np.zeros(n_assets, dtype=np.bool_)

valid_mask = np.zeros(n_assets, dtype=bool)

Copilot · 2025-11-15T14:26:48Z

src/quant_research_starter/benchmarks/bench_opt.py

+    return elapsed, results
+
+
+def benchmark_numba(prices: pd.DataFrame, signals: pd.DataFrame) -> float:


The return type hint for benchmark_numba is incorrect. The function returns a tuple (elapsed, results) or (None, None) but is annotated to return float. The return type should be tuple[float | None, dict | None] or similar.

Suggested change

def benchmark_numba(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

def benchmark_numba(prices: pd.DataFrame, signals: pd.DataFrame) -> tuple[float | None, dict | None]:

Copilot · 2025-11-15T14:26:49Z

src/quant_research_starter/benchmarks/bench_opt.py

+def compute_rank_weights_numba(signals, max_leverage, long_pct, short_pct):
+    """Helper to compute rank weights using Numba."""
+    return rank_based_weights(signals, max_leverage, long_pct, short_pct)


The helper function compute_rank_weights_numba is redundant - it's a trivial wrapper that simply calls rank_based_weights with the same parameters. Consider calling rank_based_weights directly on line 80 to reduce unnecessary indirection.

Copilot · 2025-11-15T14:26:49Z

src/quant_research_starter/benchmarks/bench_opt.py

+    returns_df = prices.pct_change().dropna()
+    aligned_signals = signals.loc[returns_df.index]
+
+    returns_arr = returns_df.values
+    n_days, n_assets = returns_arr.shape
+
+    weights_list = []
+    current_weights = np.zeros(n_assets, dtype=np.float64)
+    for date in returns_df.index:
+        signal_row = aligned_signals.loc[date].values.astype(np.float64)
+        weights = compute_rank_weights_numba(signal_row, 1.0, 0.9, 0.1)
+        current_weights = weights.copy()
+        weights_list.append(current_weights)
+
+    weights = np.array(weights_list, dtype=np.float64)
+    weights_prev = np.vstack([np.zeros((1, n_assets), dtype=np.float64), weights[:-1]])
+
+    turnover = compute_turnover(weights, weights_prev)
+    strat_ret = compute_strategy_returns(
+        weights_prev, returns_arr.astype(np.float64), turnover, 0.001
+    )
+    portfolio_value = compute_portfolio_value(strat_ret, 1_000_000.0)
+
+    elapsed = time.perf_counter() - start
+    results = {
+        "portfolio_value": pd.Series(portfolio_value, index=returns_df.index),
+        "returns": pd.Series(
+            np.diff(portfolio_value) / portfolio_value[:-1], index=returns_df.index[1:]
+        ),
+    }
+    return elapsed, results
+
+
+def compute_rank_weights_numba(signals, max_leverage, long_pct, short_pct):
+    """Helper to compute rank weights using Numba."""
+    return rank_based_weights(signals, max_leverage, long_pct, short_pct)
+
+
+def benchmark_cython(prices: pd.DataFrame, signals: pd.DataFrame) -> float:
+    """Benchmark Cython-accelerated implementation."""
+    if not CYTHON_AVAILABLE:
+        return None, None
+
+    start = time.perf_counter()
+
+    returns_df = prices.pct_change().dropna()
+    aligned_signals = signals.loc[returns_df.index]
+
+    returns_arr = returns_df.values.astype(np.float64)
+    n_days, n_assets = returns_arr.shape
+
+    weights_list = []
+    current_weights = np.zeros(n_assets, dtype=np.float64)
+    for date in returns_df.index:
+        signal_row = aligned_signals.loc[date].values.astype(np.float64)
+        weights = compute_rank_weights_cython(signal_row, 1.0)
+        current_weights = weights
+        weights_list.append(current_weights)
+
+    weights = np.array(weights_list, dtype=np.float64)
+    weights_prev = np.vstack([np.zeros((1, n_assets), dtype=np.float64), weights[:-1]])
+
+    turnover = compute_turnover_cython(weights, weights_prev)
+    strat_ret = compute_strategy_returns_cython(
+        weights_prev, returns_arr, turnover, 0.001
+    )
+
+    portfolio_value = compute_portfolio_value(strat_ret, 1_000_000)
+
+    elapsed = time.perf_counter() - start
+    results = {
+        "portfolio_value": pd.Series(portfolio_value, index=returns_df.index),
+        "returns": pd.Series(
+            np.diff(portfolio_value) / portfolio_value[:-1], index=returns_df.index[1:]
+        ),
+    }
+    return elapsed, results
+
+
+def compute_rank_weights_cython(signals, max_leverage):
+    """Helper to compute rank weights (simplified for Cython benchmark)."""
+    valid_mask = ~np.isnan(signals)
+    valid_signals = signals[valid_mask]
+    if len(valid_signals) == 0:
+        return np.zeros_like(signals)
+
+    ranks = np.argsort(np.argsort(valid_signals)) + 1
+    long_threshold = np.percentile(ranks, 90)
+    short_threshold = np.percentile(ranks, 10)
+
+    weights = np.zeros_like(signals)
+    valid_idx = 0
+    for i in range(len(signals)):
+        if valid_mask[i]:
+            if ranks[valid_idx] >= long_threshold:
+                weights[i] = 1.0
+            elif ranks[valid_idx] <= short_threshold:
+                weights[i] = -1.0
+            valid_idx += 1
+
+    long_count = (weights > 0).sum()
+    short_count = (weights < 0).sum()
+
+    if long_count > 0:
+        weights[weights > 0] = 1.0 / long_count
+    if short_count > 0:
+        weights[weights < 0] = -1.0 / short_count
+
+    total_leverage = abs(weights).sum()
+    if total_leverage > max_leverage:
+        weights *= max_leverage / total_leverage
+
+    return weights


There is significant code duplication between benchmark_numba (lines 70-91), benchmark_cython (lines 115-137), and partially in compute_rank_weights_cython (lines 151-182). The data preprocessing logic (computing returns, aligning signals, building weights arrays) is repeated. Consider extracting this common logic into a shared helper function to improve maintainability.

Copilot · 2025-11-15T14:26:49Z

src/quant_research_starter/benchmarks/bench_opt.py

+    n_days, n_assets = returns_arr.shape
+
+    weights_list = []
+    current_weights = np.zeros(n_assets, dtype=np.float64)


This assignment to 'current_weights' is unnecessary as it is redefined before this value is used.

Suggested change

current_weights = np.zeros(n_assets, dtype=np.float64)

# current_weights = np.zeros(n_assets, dtype=np.float64) # Removed unnecessary assignment

feat: numba

818058b

ayushkrtiwari added Semver:minor minor version changes Type:Hard senior developers, max points labels Nov 15, 2025

ayushkrtiwari requested a review from Copilot November 15, 2025 14:21

Copilot started reviewing on behalf of ayushkrtiwari November 15, 2025 14:21 View session

Copilot finished reviewing on behalf of ayushkrtiwari November 15, 2025 14:24

ayushkrtiwari merged commit a27eeda into OPCODE-Open-Spring-Fest:main Nov 15, 2025
17 of 20 checks passed

Copilot AI reviewed Nov 15, 2025

View reviewed changes

		return prices, signals


		def benchmark_vanilla(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

		return rank_based_weights(signals, max_leverage, long_pct, short_pct)


		def benchmark_cython(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

	valid_mask = np.zeros(n_assets, dtype=np.bool_)
	valid_mask = np.zeros(n_assets, dtype=bool)

		return elapsed, results


		def benchmark_numba(prices: pd.DataFrame, signals: pd.DataFrame) -> float:

	current_weights = np.zeros(n_assets, dtype=np.float64)
	# current_weights = np.zeros(n_assets, dtype=np.float64) # Removed unnecessary assignment

feat: numba #143

feat: numba #143

Uh oh!

Conversation

Satvik-Singh192 commented Nov 15, 2025

Description

Key Enhancements

Semver Changes

Issues

Checklist

Uh oh!

netlify bot commented Nov 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

✅ Deploy Preview for strong-duckanoo-898b2c ready!

Uh oh!

Satvik-Singh192 commented Nov 15, 2025

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Nov 15, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

netlify bot commented Nov 15, 2025 •

edited

Loading