`CumSum`#

Description#

The CumSum function returns the running sum of all samples seen since the start of the stream (or since the last reset). It is the streaming equivalent of numpy.cumsum. Memory is O(1) regardless of how many samples have been processed.

Equation:

\[ y[t] = \sum_{i=0}^{t} x[i] \]

Parameters: none.

NaN handling: NaN propagates by ordinary IEEE-754 addition. Once a NaN enters the input, every subsequent output is NaN. This matches numpy.cumsum, not pandas.Series.cumsum(skipna=True).

NaN handling#

Policy: ignore. A NaN in any input at index t causes the function to skip that step: output at t is NaN and internal state is unchanged. Subsequent finite samples are processed as if step t had not occurred.

Examples#

Usage example#

import numpy as np
import plotly.graph_objects as go
from plotly.subplots import make_subplots
from screamer import CumSum

data = np.random.default_rng(0).normal(0.05, 1.0, size=100)
cumulative = CumSum()(data)

fig = make_subplots(rows=2, cols=1, shared_xaxes=True,
                    row_heights=[0.5, 0.5], vertical_spacing=0.1)
fig.add_trace(go.Scatter(y=data, mode='lines', name='Daily return'), row=1, col=1)
fig.add_trace(go.Scatter(y=cumulative, mode='lines',
                         name='Cumulative return',
                         line=dict(color='green')), row=2, col=1)
fig.update_layout(
    title="CumSum: Running Total of a Random Walk Increment",
    xaxis_title="Index",
    yaxis_title="x[t]",
    yaxis2_title="CumSum(x)[t]",
    margin=dict(l=20, r=20, t=60, b=20),
    legend=dict(orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1)
)
fig.show()

Implementation Details#

CumSum keeps a single double accumulator. Every input is added in-place, and the running total is returned. There is no warmup. The numpy reference is numpy.cumsum.

CumSum#