Engine comparison · v0.2 · 50-strategy benchmark

PineForge vs PyneCore.
Reproducible, not rhetorical.

Every number on this page is generated by bash benchmarks/run_all.sh in the open-source pineforge-engine repo, against the same 41,307-bar Binance ETH/USDT 15-minute feed. Reproduce in ~3 minutes from a clean clone, zero external API calls.

50-strategy match degree

How many of 50 strategies hit excellent tier against TradingView.

C++ static lib

PineForge

49 / 50

Excellent49Strong1Moderate0Weak0

Python (PyneSys cloud-compiled)

PyneCore

46 / 50

Excellent46Strong1Moderate2Weak1

TypeScript (LuxAlgo)

PineTS

indicators only

Strategy backtester—Per-bar indicators✓10/10 indicatorsmatch

Strategy execution is on the PineTS roadmap. We benchmark indicator-precision against PineTS to triangulate floating-point divergences.

Tiers follow the canonical PineForge parity sweep: excellent = all four dimensions (count delta, entry p90, exit p90, P&L p90) within strict thresholds and ≥95% trades matched; strong within 5× strict; moderate / weak / minimal step down from there. Strategies that use TradingView’s trail_* exits get the production threshold profile (looser exit + P&L tolerances).

The 3-strategy delta

Three strategies draw all the daylight.

On 47 of 50 reference strategies PineForge and PyneCore both hit excellent. The 3-strategy gap is not random — every divergence is in the same category: bracket exits, trailing stops, or partial position closes. PyneCore’s broker emulator differs from TV here; PineForge mirrors TV trade-for-trade.

06-liquidity-sweep

bracket exit

PineForgeexcellent (88 / 88)·PyneCoremoderate (91)

93 TV trades in window. PineForge matches 88 within strict tolerances. PyneCore generates 91 trades — a +3 count drift, plus exit-price drift on bracket-stopped exits.

07-scalping-strategy

trailing stop (production thresholds)

PineForgeexcellent (412 / 429)·PyneCoremoderate (412)

429 TV trades in window. PineForge: 412 matched, all four parity dimensions inside production thresholds. PyneCore: same matched count but exit-price p90 outside threshold — broker-emulator trail_offset arithmetic diverges from TV.

49-partial-exit-qty-percent

partial close (qty_percent)

PineForgeexcellent (683 / 725)·PyneCoreweak (2,671)

The clearest divergence in the corpus. 725 TV trades, PineForge matches 683 at strict parity. PyneCore generates 2,671 trades — 3.7× the correct count. Root cause: strategy.close(qty_percent=…) in PyneCore splits each entry into per-percentage sub-exits instead of a single partial close. Open upstream issue as of this commit.

Architecture & capability matrix

What each engine actually is.

Capability	PineForge	PyneCore	PineTS
Backend language	C++17 static lib	Python (cloud-transpiled)	TypeScript
Distribution	open-source runtime + closed transpiler	open-source	open-source
50-strategy benchmark	49 excellent + 1 strong	46 excellent + 1 strong + 2 moderate + 1 weak	— (not in scope)
158/162 strict TV parity	✓	— (different corpus)	— (indicators only)
Determinism guarantee	byte-reproducible	best-effort	best-effort
Bracket / trail / partial-exit semantics	TV-faithful	drift documented	N/A
Bar magnifier (intra-bar fills)	6 distributions	subset	N/A
Forward-test engine	Q3 2026	today	N/A
Live broker integration	2027	today (subset)	N/A
Hosted Studio UI	Q4 2026	CLI only	CLI only
Strategy marketplace + DRM	2027 (designed)	—	—

Where each engine wins

We don’t hide our gaps. Neither should they.

CHOOSE PINEFORGE WHEN

You need byte-reproducible determinism (CI gates, audit trails, paid-parity claims to clients).
You need TV-faithful semantics on bracket exits, trailing stops, or partial closes. Three concrete strategies above are unambiguous on this.
You need native compiled speed for parameter sweeps (Optuna across thousands of parameter combinations on 50k-bar feeds).
You want a hosted Studio UI later — Code · Backtest · Optimize · Compare · Reports tabs are coming Q4 2026.
You eventually want to sell compiled strategies to other traders. The encrypted-distribution + license-server design is in the public engine repo.

CHOOSE PYNECORE WHEN

You need forward-testing or live broker execution today. PineForge ships those Q3-Q4 2026; PyneCore has them now.
You need a fully-Python strategy execution path (deeper integration with NumPy/Pandas backtesting tooling, Jupyter-native iteration).
You’re comfortable on the bracket/trail/partial-exit caveats (47/50 of strategies don’t exercise them).
The fully open-source ethos matters more than the closed transpiler tradeoff. PyneCore is open end-to-end; PineForge’s runtime is OSS but the codegen is closed.
You’re a heavy contributor and want a project where your PRs land directly in the strategy execution path.

Indicator precision · 10 canonical TA functions

PineForge & PineTS agree to 10⁻¹⁰.
PyneCore drifts at 10⁻⁸.

On the 41,307-bar feed, PineForge and PineTS produce per-bar relative errors at the floating-point limit (≈10⁻¹⁰ p90). PineForge↔PyneCore drifts ~100× higher (~10⁻⁸ p90). Both are well below trade-execution tolerance — but the divergence is consistent, suggesting a systematic rounding-mode difference rather than randomness.

Indicator	PF↔PineTS p90	PF↔PyneCore p90
ema21	1.9e-10	1.9e-08
sma21	1.9e-10	1.9e-08
rsi14	9.7e-11	9.7e-09
atr14	2.8e-10	2.8e-08
macd_line	2.3e-10	2.3e-08
macd_signal	2.4e-10	2.4e-08
bb_basis	0	0
bb_upper	1.9e-10	1.9e-08

Don’t trust the table. Reproduce it.

Every number on this page is generated by the public benchmark suite. No hidden config, no API keys, no committed-snapshot tricks. ~3 minutes from a clean clone.

# 1. Clone the open-source engine + benchmark suite
git clone https://github.com/fullpass-4pass/pineforge-engine
cd pineforge-engine

# 2. Pull the LFS-tracked OHLCV (2.3 MB)
git lfs install && git lfs pull

# 3. Run the full three-engine sweep (~3 min)
bash benchmarks/run_all.sh

# 4. Read the results — same table that's on this page
cat benchmarks/results/summary.md

See benchmarks repo →Join waitlist

PineForge vs PyneCore.Reproducible, not rhetorical.