An artifact is not a publication. It needs a token to download, expires after ninety days, and nothing outside GitHub can link to it, so a number that only lives in an artifact is a number nobody can check. A third job merges a green run's two payloads into benchmarks/vs_vectorbt/results/latest.json and commits it. The two are stored side by side rather than folded into one table: they run on two runners, and timings from two machines are not rows of the same table. Only a run where both measuring jobs came back green is published, and a dispatch that pins an old version measures and reports without becoming the published number.
manifoldbt vs vectorbt vs raptorbt
An engine-to-engine benchmark you can re-run yourself. It installs every engine from PyPI, generates its own data, checks that they produced the same result, and only then reports how long each took.
pip install manifoldbt
pip install -r requirements-lock.txt
python bench.py --bars 10000 100000 1000000 --reps 7 --cold-start-reps 3 --memory-bars 2000000 --out results.json
python report.py results.json
No dataset to download, no API key, no configuration. The same command runs in GitHub Actions on a public runner, so every published number has a run URL behind it.
Python 3.12, and not by preference. raptorbt is built against pyo3 0.20.3,
whose maximum supported CPython is 3.12: no release up to 0.9.0 publishes a
cp313 wheel, and a source build refuses outright. Comparing engines means
running them in one environment, and the environment has to be one all of them
support. Add or drop a challenger with --engines; an engine that is not
installed is skipped with a printed line rather than crashing the run.
manifoldbt is the reference. Every parity check and every ratio is a challenger against it, never two challengers against each other: three engines make three pairs, and a table of pairs is a matrix, not a benchmark. The reference gets no advantage from the position, it is simply the one every timing is divided by.
The rule this harness is built around
A speed comparison between backtesters is worthless unless they did the same work. So parity comes first:
- each workload runs once per engine;
- total return, round-trip count and fees are compared against the reference;
- a workload an engine disagrees on gets no published timing for it.
Three verdicts come out of that gate:
| Verdict | Meaning | What gets published |
|---|---|---|
exact |
agreement down to float-reordering noise (relative tolerance 1e-9) | the timing, in the headline table |
documented |
that engine disagrees, the workload declared it in advance for that engine, and the cause is written down | the timing, in an annex, with the cause and its measured size |
failed |
it disagrees and nobody predicted it | nothing. The run exits non-zero |
A fourth outcome sits beside the gate rather than inside it. unsupported
means an engine cannot express the workload at all: it is not run, and the
report prints the reason instead of an empty cell. A blank in a speed table
reads as a defeat, and "this engine has no fixed-quantity sizing" is not a
defeat, it is a different fact.
The failed path is not decoration. It is the reason the other numbers can be
trusted, and it makes the benchmark fail loudly if a future release of either
engine changes a fill rule.
Method
Interleaved repetitions. The engines alternate inside each repetition (A, B, A, B, ...) rather than running in two blocks. On a shared cloud runner that slows down halfway through, two blocks would hand the penalty to whichever engine ran second; alternating splits it evenly.
The ratio is the headline, milliseconds are context. Each ratio comes from two measurements taken seconds apart on the same machine. Absolute timings from a shared runner are worth much less than the ratio between them.
Dispersion is published, and noise is flagged. Every point carries min,
median, max and interquartile range. If the IQR exceeds 15% of the median the
point is marked noisy and is not headline material, however good it looks.
Warmup is discarded, and that favours vectorbt on purpose. The first call of each engine is thrown away, which is where vectorbt pays its numba compilation. Charging a one-off JIT cost to every repetition would inflate the result.
Data loading is excluded on both sides. manifoldbt is handed a prepared store, vectorbt is handed prepared Series. What is timed is indicators plus simulation plus reading the headline metrics, nothing else.
Only public APIs. manifoldbt is driven through bt.run(strategy, config, store), the documented entry point, not through an internal fast path.
What is compared
| Workload | What it exercises | vectorbt | raptorbt |
|---|---|---|---|
sma_cross |
SMA 30/150 crossover, long-only, no cost | exact | exact |
sma_cross_metrics |
the same simulation, plus max drawdown, Sharpe, Sortino and volatility | exact | exact |
sma_cross_costs |
the same signal with a 5 bps fee and 2 bps of slippage | exact | unsupported |
multi_asset |
five assets in one shared book, capped at 1M bars | exact | unsupported |
ema_rsi_fees |
EMA 12/26 with an RSI(14) filter and a 5 bps taker fee, capped at 1M bars | exact | unsupported |
bracket_sl_tp |
the same entry with a 15 bps stop and a 30 bps target | documented | documented |
Costs live in their own workload, and not by preference
A cost model cannot simply be switched on across the board. Measured on the
headline workload, a fee or a slippage applied to FractionOfEquity sizing puts
the engines 1.3e-4 and 2.1e-5 of capital apart against a 1e-9 tolerance, while
the round-trip counts stay identical: the trading agrees, the cost arithmetic
does not, because one charges the fee on top of the notional and the other
reserves it out of cash first. In fixed units both land exactly, to 1e-15.
So sma_cross_costs carries both costs on the same signal as the headline, and
the price of that is visible: x48.0 against x50.6 at 100k bars, x36.1 against
x36.7 at 1M. Costs change the result, not the ratio.
The multi-asset workload is the one manifoldbt does worst on
It is here on purpose. A portfolio is what people actually run, and it is also where the two designs differ in kind: manifoldbt walks a universe against one shared book, while vectorbt broadcasts a column per asset and has to be told to share cash at all.
Broadcasting wins.
| Workload | 100k bars | 1M bars |
|---|---|---|
sma_cross, one asset |
x50.6 | x36.7 |
multi_asset, five assets |
x10.7 | x8.8 |
The absolute timings say why: going from one asset to five costs manifoldbt 6.1x (26.4 ms to 162.3 ms at 1M bars) and vectorbt 1.4x (989 ms to 1415 ms). Five columns in one vectorised pass is close to free for it; five books are not free for anything that walks them.
Publishing this lowers the average number on the page. It is the most useful workload in the suite for anyone deciding whether to switch, which is the only audience the benchmark has.
Why the fee workload stops at 1M bars
A workload can stop being a comparison before it stops running. ema_rsi_fees
sizes in fixed units and pays 5 bps a side, and at 1-minute resolution it turns
over often enough that the fees compound into the account: measured, it ends at
-15% of capital on 1M bars, -74% on 5M, and exactly -100% on 10M, where fees
reach 99,611 of the 100,000 it started with.
Past that point the engines still agree on the equity, because both are sitting at zero, and disagree by thousands of round-trips about how many more worthless trades to book on a dead account. That is a fact about a bankrupt strategy, not about either engine, so the workload carries a ceiling and the runner skips it above that with the reason printed. The other workloads have no ceiling.
What the windows are, and why they moved
sma_cross crosses on 30/150 rather than 10/50. The two were measured against
each other on the same 5M bars, and the slower pair is worth about 15% on the
ratio (x267 against x232 with a performance summary) because it books a third of
the trades and manifoldbt's cost, unlike vectorbt's, moves with the trade count.
That is a real effect and a small one, and it is worth knowing which way the knobs turn before anyone quotes a number:
| Turn up | Effect on the ratio | Why |
|---|---|---|
| series length | widens | vectorbt materialises the simulation; its cost is linear in bars |
| asking for the summary | widens sharply | it has to build the equity curve it deferred |
| number of trades | narrows | near-free for vectorbt's per-bar loop, real for manifoldbt |
| number of indicators | narrows | same reason |
Measured at 5M bars on four levels of turnover, the ratio runs from x40 at 480,000 round-trips to x71 at 2,500. The floor matters more than the peak: even in the busiest configuration tested, with half a million round-trips, the gap holds at x40, and x151 with a performance summary.
Each of those runs across a range of series lengths. Two further axes, cold start and memory, are measured in their own processes because they cannot be measured honestly inside the main one.
Scope, stated twice on purpose. sma_cross and sma_cross_metrics run the
identical simulation; only the second one also produces a performance summary.
manifoldbt and raptorbt both compute that summary inside the run whether or not
you read it, while vectorbt defers the equity curve until a risk metric asks for
it and then pays to materialise it. Reporting both scopes is the only honest way
to present the result: a reader who only wants a total return should look at the
first number, and a reader who wants a Sharpe should look at the second. Parity
on the summary workload is gated on total return, round-trip count and max
drawdown, which match exactly across all three; the vectorbt ratios agree to
about 3e-4, because manifoldbt buckets its daily returns slightly differently,
and that residual is reported rather than smoothed over.
raptorbt annualises its ratios on its own basis and moves that basis between releases (the same run reads Sharpe 0.21 on 0.4.1 and 3.43 on 0.9.0, against manifoldbt's 8.14), so its Sharpe and Sortino are recorded but not compared: subtracting them from the reference would publish a units mismatch as a disagreement. It reports no volatility at all, and the harness leaves that cell empty rather than recomputing it from the equity curve, which would credit raptorbt with the harness's own arithmetic. None of this touches the gate, which runs on money, round-trips and drawdown, all three basis-free. Its drawdown matches the reference to 4e-15.
The vectorbt side of the summary is written out in pandas rather than through
pf.sharpe_ratio() for two reasons, both in vectorbt's favour or neutral.
manifoldbt computes its ratios on daily returns annualised by sqrt(365) and its
drawdown at full bar resolution, so the native accessors would return different
numbers and the comparison would be timing two different computations; and the
hand-written version is measurably faster than a single native accessor on the
same data, so vectorbt is credited with the quicker of its two paths.
Cold start. The steady-state table discards a warmup call, which is where vectorbt compiles its numba kernels. A user pays that cost in every new notebook, script or CI job, so it is measured rather than waved away: a fresh process, an engine it has never imported, one backtest. The Python, numpy and pandas baseline is measured the same way and reported alongside, so the engine's own share can be read off instead of argued about.
Memory added by the run. Resident memory sampled while the backtest executes, after a warmup. vectorbt materialises the simulation as arrays and its footprint grows with the series; manifoldbt streams bars out of its store. What is compared is what running a backtest costs on top of already holding the data: building each engine's data representation is a different question and is excluded on both sides, in manifoldbt's case a deliberately unflattering choice since its one-off ingest is the memory-hungry part.
Why raptorbt sits out the fee workload
Two blockers, either one sufficient, both measured rather than assumed.
Sizing. raptorbt has no fixed-quantity mode. position_sizes is a fraction of
equity (a constant 0.5 buys exactly half the equity of the bar before the entry),
lot_size rounds a computed size down to a multiple of itself, and
alloted_capital fixes the notional rather than the quantity. Reproducing
units=5 would mean feeding a fraction derived from an equity curve that does
not exist until the run is over.
Indicator. raptorbt.ema seeds on a simple mean of the first period bars and
emits from bar period-1; manifoldbt seeds on the first observation and emits
from bar 0. Same recursion, different warmup, so the signal differs early and the
round-trip count with it. Its sma matches the reference to 3.3e-13 and its
rsi matches to the last bit, which is why the other three workloads run.
Running it anyway with a different size and a different indicator would produce a number, and the number would not mean anything.
Why the fee workload sizes in units
With FractionOfEquity sizing and a non-zero fee the engines size differently:
manifoldbt charges the fee on top of a full-equity notional, vectorbt reserves
it out of cash first. Both are legitimate product decisions, and comparing them
would compare policy rather than speed or correctness. Sizing in fixed units
isolates the fee arithmetic, which is the thing both engines must agree on.
The documented divergence, in full
When a bracket fires intrabar and the entry condition still holds at that bar's close, the three engines take three different roads:
- manifoldbt books two orders on that bar: the stop or target exit, then a fresh entry at the close.
- vectorbt processes one order per bar and re-enters on the next bar.
- raptorbt does not re-arm at all. The level still being true is not enough; it waits for the level to go false and true again.
None of them is wrong. On controlled bars the bracket fills themselves match exactly, which the cross-engine parity suite shipped with the library pins test by test; the divergence is purely about when a re-entry is allowed.
The harness counts the affected round-trips rather than hand-waving at them, and it is one population, not three: at 10,000 bars manifoldbt books 214 round-trips of which 82 re-enter on the exit bar, and raptorbt books exactly 214 - 82 = 132. The report states the count and the share, so the size of the difference is measured instead of asserted.
Alignment choices, and why each one exists
Engines only produce identical numbers if they are told to do the same thing. These are the conventions the harness sets, all of them visible in engine_mbt.py, engine_vbt.py and engine_rbt.py:
signal_delay=0withexecution_price="AtClose", matching whatPortfolio.from_signalsdoes by default: a signal fills at the close of the bar that produced it.warmup_bars=0, so the indicator's own NaN warmup is what suppresses early signals, identically on both sides.- Signals are fed to vectorbt as a level (
entries= the condition holds,exits= it no longer holds) rather than as transitions, which reproduces manifoldbt's target-position semantics. - Indicators are mirrored on manifoldbt's exact definitions. The EMA seeds on
the first observation with
alpha = 2/(span+1), whichewm(span=n, adjust=False)reproduces exactly. The RSI is Wilder's, seeded with the simple average of the firstperioddeltas and emitted from barperiod: a plainewm(alpha=1/period)over the whole delta series is a different indicator, and using it would have made the engines disagree for a reason that has nothing to do with either engine. - The generated bars have no opening gap (
open == previous close). A bar that gaps through a stop is the one case where two engines can legitimately book different fill prices while both being correct, so that class of false failures is removed from the data rather than argued about in the report.
On the raptorbt side specifically:
upon_bar_close=True, which fills at the close of the signal bar. Turning it off does not add a one-bar delay, it moves the fill to that same bar's open, so there is no setting on that side matching a delayed execution.- Sizing is left at its default, which takes the whole equity of the bar before
the entry. Since the account is flat at that point, that is the same number as
the equity at the fill, and the three engines size identically:
sma_crosscomes back with manifoldbt's final equity to the last bit. - The bracket is set on the config, in fractions rather than percent: 0.15 means
a 15% stop, so the workload's 0.15% is
0.0015. Passing the percent number is not an error, it is a stop so wide it never triggers. - Indicators come from raptorbt's own Rust
smaandrsi, not from numpy: that is what a user would write, and it is what deserves to be timed.
Reading the numbers honestly
- On a shared runner with 4 vCPUs, an engine that parallelises is understated. These are floors, not peaks.
- The published wheels are CPU-only, and GitHub-hosted runners have no GPU, so nothing here says anything about GPU performance.
- vectorbt is the open-source package (
pip install vectorbt), not vectorbtpro. - raptorbt has no fan-out API for a parameter grid on one instrument, so its
sweep column is a Python loop over
run_single_backtest. That is not a handicap the harness imposed, it is the only spelling available, and the moving averages are hoisted out of the loop so it gets the same courtesy vectorbt gets on its own grid path.
Files
data.py- deterministic OHLCV generator and its content digestprobe_child.py- the cold-start and memory probes, each in a fresh processworkloads.py- the parameters every engine reads, and the per-engine notesengines.py- the registry: who is in the comparison, and who is the referenceengine_mbt.py/engine_vbt.py/engine_rbt.py- one adapter per engineparity.py- the gatebench.py- the runnersweep_child.py- one parameter-grid point, in its own processreport.py- JSON to Markdown, and to the GitHub job summaryci_activate.py- activates the licence the sweep job needs, and refuses to continue without oneci/bench-vs-vectorbt.yml- the workflow, deployed to the public repository
Parameter sweeps
Sweeps run in their own CI job, because they need a licence: an unlicensed fan-out call waits out a fixed interval before doing any work, so a stopwatch would be timing the wait. The harness refuses to produce a number in that state rather than producing a wrong one.
Three points, sized from what the runner actually did rather than guessed. Measured there: 87.5 us per combination for manifoldbt at 20,000 bars, 1.16 ms for vectorbt, 1.34 ms for raptorbt, and 15.0 ms for raptorbt at 200,000 bars.
Only the first point is one vectorbt can hold. It materialises the simulation per combination -- 1.57 MB of it at 20,000 bars -- so 5,000 combinations already cost it 2.5 GB and 20,000 would need 31 GB. The other two are out of its scope, and that is where a sweep stops being a speed comparison and becomes a capability one: the question is no longer how long it takes but whether the machine can hold it at all.
raptorbt sets the time budget rather than manifoldbt. With no fan-out API for a parameter grid on one instrument, its sweep is a Python loop costing a full backtest per cell, which is why the large point goes deep in combinations on a short series instead of the reverse: 20,000 combinations on 20,000 bars costs it 27 seconds a call, where 5,000 combinations on a million bars would cost it 25 minutes.
Sweep ratios depend on the machine far more than single-backtest ratios do, because manifoldbt is the only one of the three that fans out across cores. The same point measured x13 on a four-vCPU cloud runner and x32 on two fast desktop cores; the CI numbers are the conservative ones, and they are the ones published.