Performance¶

Accumulators¶

Most rolling indicators implemented with Pandas and NumPy require running over data more than once. For example, a simple sum of three arrays involves at least two passes over data. Moreover, if you want to calculate such an indicator iterativelly (i.e., bar by bar), you either need to pre-calculate it entirely and store in memory, or re-calculate each window, which may dramatically hit performance. Accumulators, on the other hand, keep an internal state that allows you to calculate an indicator value every time a new data point arrives, leading to the best performance possible.

Design a one-pass rolling z-score

>>> @njit
... def fastest_rolling_zscore_1d_nb(arr, window, minp=None, ddof=1):
...     if minp is None:
...         minp = window
...     out = np.full(arr.shape, np.nan)
...     cumsum = 0.0
...     cumsum_sq = 0.0
...     nancnt = 0
...     
...     for i in range(len(arr)):
...         pre_window_value = arr[i - window] if i - window >= 0 else np.nan
...         mean_in_state = vbt.nb.RollMeanAIS(
...             i, arr[i], pre_window_value, cumsum, nancnt, window, minp
...         )
...         mean_out_state = vbt.nb.rolling_mean_acc_nb(mean_in_state)
...         _, _, _, mean = mean_out_state
...         std_in_state = vbt.nb.RollStdAIS(
...             i, arr[i], pre_window_value, cumsum, cumsum_sq, nancnt, window, minp, ddof
...         )
...         std_out_state = vbt.nb.rolling_std_acc_nb(std_in_state)
...         cumsum, cumsum_sq, nancnt, _, std = std_out_state
...         out[i] = (arr[i] - mean) / std
...     return out

>>> data = vbt.YFData.pull("BTC-USD")
>>> rolling_zscore = fastest_rolling_zscore_1d_nb(data.returns.values, 14)
>>> data.symbol_wrapper.wrap(rolling_zscore)
Date
2014-09-17 00:00:00+00:00         NaN
2014-09-18 00:00:00+00:00         NaN
2014-09-19 00:00:00+00:00         NaN
                                  ...   
2023-02-01 00:00:00+00:00    0.582381
2023-02-02 00:00:00+00:00   -0.705441
2023-02-03 00:00:00+00:00   -0.217880
Freq: D, Name: BTC-USD, Length: 3062, dtype: float64

>>> (data.returns - data.returns.rolling(14).mean()) / data.returns.rolling(14).std()
Date
2014-09-17 00:00:00+00:00         NaN
2014-09-18 00:00:00+00:00         NaN
2014-09-19 00:00:00+00:00         NaN
                                  ...   
2023-02-01 00:00:00+00:00    0.582381
2023-02-02 00:00:00+00:00   -0.705441
2023-02-03 00:00:00+00:00   -0.217880
Freq: D, Name: Close, Length: 3062, dtype: float64

Chunking¶

An innovative new chunking mechanism that takes a specification of how arguments should be chunked, automatically splits array-like arguments, passes each chunk to the function for execution, and merges back the results. This way, you can split large arrays and run any function in a distributed manner! Additionally, VBT implements a central registry and provides the chunking specification for all arguments of most Numba-compiled functions, including the simulation functions. Chunking can be enabled by a single command. No more out-of-memory errors!

Backtest at most 100 parameter combinations at once

>>> @vbt.chunked(
...     chunk_len=100,
...     merge_func="concat",  # (1)!
...     execute_kwargs=dict(  # (2)!
...         show_progress=True,
...         clear_cache=True,
...         collect_garbage=True
...     )
... )
... def backtest(data, fast_windows, slow_windows):  # (3)!
...     fast_ma = vbt.MA.run(data.close, fast_windows, short_name="fast")
...     slow_ma = vbt.MA.run(data.close, slow_windows, short_name="slow")
...     entries = fast_ma.ma_crossed_above(slow_ma)
...     exits = fast_ma.ma_crossed_below(slow_ma)
...     pf = vbt.PF.from_signals(data.close, entries, exits)
...     return pf.total_return

>>> param_product = vbt.combine_params(  # (4)!
...     dict(
...         fast_window=vbt.Param(range(2, 100), condition="fast_window < slow_window"),
...         slow_window=vbt.Param(range(2, 100)),
...     ),
...     build_index=False
... )
>>> backtest(
...     vbt.YFData.pull(["BTC-USD", "ETH-USD"]),  # (5)!
...     vbt.Chunked(param_product["fast_window"]),  # (6)!
...     vbt.Chunked(param_product["slow_window"])
... )

Concatenate the Series returned by each chunk into one Series
Show a progress bar, but also clear cache and collect garbage after processing each chunk
Function that takes a data instance but also two parameter arrays: fast and slow window lengths. Both arrays will have the same number of values; for example, the first combination corresponds to the first value in fast_windows and the first value in slow_windows.
Generate conditional parameter combinations
Don't split data into chunks
Split both parameter arrays into chunks

Chunk 48/48

fast_window  slow_window  symbol 
2            3            BTC-USD    193.124482
                          ETH-USD     12.247315
             4            BTC-USD    159.600953
                          ETH-USD     15.825041
             5            BTC-USD    124.703676
                                        ...    
97           98           ETH-USD      3.947346
             99           BTC-USD     25.551881
                          ETH-USD      3.442949
98           99           BTC-USD     27.943574
                          ETH-USD      3.540720
Name: total_return, Length: 9506, dtype: float64

Parallel Numba¶

Most Numba-compiled functions were rewritten to process columns in parallel using automatic parallelization with @jit, which can be enabled by a single command. Best suited for lightweight functions applied on wide arrays.

Benchmark the rolling mean without and with parallelization

>>> df = pd.DataFrame(np.random.uniform(size=(1000, 1000)))

>>> %timeit df.rolling(10).mean()  # (1)!
45.6 ms ± 138 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

>>> %timeit df.vbt.rolling_mean(10)  # (2)!
5.33 ms ± 302 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit df.vbt.rolling_mean(10, jitted=dict(parallel=True))  # (3)!
1.82 ms ± 5.21 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Using Pandas
Using Numba without parallelization
Using Numba with parallelization

Multithreading¶

Integration of ThreadPoolExecutor from concurrent.futures, ThreadPool from pathos, and Dask backend for running multiple chunks across multiple threads. Best suited for accelerating heavyweight functions that release GIL, such as Numba and C functions. Multithreading + Chunking + Numba =

Benchmark 1000 random portfolios without and with multithreading

>>> data = vbt.YFData.pull(["BTC-USD", "ETH-USD"])

>>> %timeit vbt.PF.from_random_signals(data.close, n=[100] * 1000)
613 ms ± 37.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit vbt.PF.from_random_signals(data.close, n=[100] * 1000, chunked="threadpool")
294 ms ± 8.91 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Multiprocessing¶

Integration of ProcessPoolExecutor from concurrent.futures, ProcessPool and ParallelPool from pathos, WorkerPool from mpire, and Ray backend for running multiple chunks across multiple processes. Best suited for accelerating heavyweight functions that do not release GIL, such as regular Python functions, and accept leightweight arguments that are easy to serialize. Ever wanted to test billions of hyperparameter combinations in a matter of minutes? This is now possible by scaling functions and entire applications up in the cloud using Ray clusters

Benchmark running a slow function on each column without and with multiprocessing

>>> @vbt.chunked(
...     size=vbt.ArraySizer(arg_query="items", axis=1),
...     arg_take_spec=dict(
...         items=vbt.ArraySelector(axis=1)
...     ),
...     merge_func=np.column_stack
... )
... def bubble_sort(items):
...     items = items.copy()
...     for i in range(len(items)):
...         for j in range(len(items) - 1 - i):
...             if items[j] > items[j + 1]:
...                 items[j], items[j + 1] = items[j + 1], items[j]
...     return items

>>> items = np.random.uniform(size=(1000, 3))

>>> %timeit bubble_sort(items)
456 ms ± 1.36 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit bubble_sort(items, _execute_kwargs=dict(engine="pathos"))
165 ms ± 1.51 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Jitting¶

Jitting means just-in-time compiling. In the VBT universe though, jitting simply means accelerating. Although Numba remains the primary jitter, VBT now enables implementation of custom jitter classes, such as that for vectorized NumPy and even JAX with GPU support. Every jitted function is registered globally, so you can switch between different implementations or even disable jitting entirely using a single command.

Run different implementations of the cumulative sum

>>> data = vbt.YFData.pull("BTC-USD", start="7 days ago")
>>> log_returns = np.log1p(data.close.pct_change())
>>> log_returns.vbt.cumsum()  # (1)!
Date
2023-01-31 00:00:00+00:00    0.000000
2023-02-01 00:00:00+00:00    0.024946
2023-02-02 00:00:00+00:00    0.014271
2023-02-03 00:00:00+00:00    0.013310
2023-02-04 00:00:00+00:00    0.008288
2023-02-05 00:00:00+00:00   -0.007967
2023-02-06 00:00:00+00:00   -0.010087
Freq: D, Name: Close, dtype: float64

>>> log_returns.vbt.cumsum(jitted=False)  # (2)!
Date
2023-01-31 00:00:00+00:00    0.000000
2023-02-01 00:00:00+00:00    0.024946
2023-02-02 00:00:00+00:00    0.014271
2023-02-03 00:00:00+00:00    0.013310
2023-02-04 00:00:00+00:00    0.008288
2023-02-05 00:00:00+00:00   -0.007967
2023-02-06 00:00:00+00:00   -0.010087
Freq: D, Name: Close, dtype: float64

>>> @vbt.register_jitted(task_id_or_func=vbt.nb.nancumsum_nb)  # (3)!
... def nancumsum_np(arr):
...     return np.nancumsum(arr, axis=0)

>>> log_returns.vbt.cumsum(jitted="np")  # (4)!
Date
2023-01-31 00:00:00+00:00    0.000000
2023-02-01 00:00:00+00:00    0.024946
2023-02-02 00:00:00+00:00    0.014271
2023-02-03 00:00:00+00:00    0.013310
2023-02-04 00:00:00+00:00    0.008288
2023-02-05 00:00:00+00:00   -0.007967
2023-02-06 00:00:00+00:00   -0.010087
Freq: D, Name: Close, dtype: float64

Using the built-in Numba-compiled function
Using the built-in function but with Numba disabled → regular Python → slow!
Register a NumPy version for the built-in Numba function
Using the NumPy version

Caching¶

Caching has been reimplemented from the ground up, and now it's being managed by a central registry. This allows for tracking useful statistics of all cacheable parts of VBT, such as to display the total cached size in MB. Full control and transparency

Get the cache statistics after computing the statistics of a random portfolio

>>> data = vbt.YFData.pull("BTC-USD")
>>> pf = vbt.PF.from_random_signals(data.close, n=5)
>>> _ = pf.stats()

>>> pf.get_ca_setup().get_status_overview(
...     filter_func=lambda setup: setup.caching_enabled,
...     include=["hits", "misses", "total_size"]
... )
                                 hits  misses total_size
object                                                  
portfolio:0.drawdowns               0       1    70.9 kB
portfolio:0.exit_trades             0       1    70.5 kB
portfolio:0.filled_close            6       1    24.3 kB
portfolio:0.init_cash               3       1   32 Bytes
portfolio:0.init_position           0       1   32 Bytes
portfolio:0.init_position_value     0       1   32 Bytes
portfolio:0.init_value              5       1   32 Bytes
portfolio:0.input_value             1       1   32 Bytes
portfolio:0.orders                  9       1    69.7 kB
portfolio:0.total_profit            1       1   32 Bytes
portfolio:0.trades                  0       1    70.5 kB

Hyperfast rolling metrics¶

Rolling metrics based on returns were optimized for best performance - up to 1000x speedup!

Benchmark the rolling Sortino ratio

>>> import quantstats as qs

>>> index = pd.date_range("2020", periods=100000, freq="1min")
>>> returns = pd.Series(np.random.normal(0, 0.001, size=len(index)), index=index)

>>> %timeit qs.stats.rolling_sortino(returns, rolling_period=10)  # (1)!
2.79 s ± 24.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit returns.vbt.returns.rolling_sortino_ratio(window=10)  # (2)!
8.12 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

Using QuantStats
Using VectorBT PRO

And many more...¶

Expect more killer features to be added on a weekly basis!