RollingTSF

Description

TA-Lib's Time-Series Forecast: fits y = slope · t + intercept on the trailing window {(0, y_{t-w+1}), …, (w-1, y_t)} and returns the line evaluated at the next bar (local t = w):

\[ \text{TSF}[t] = \text{intercept} + \text{slope} \cdot w \]

1→1. Bit-exact match to talib.TSF(real, timeperiod). Composes two detail::RollingSum buffers + precomputed time-axis constants. O(1) per step.

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 RollingTSF

np.random.seed(0)
price = 100 * np.exp(np.cumsum(np.random.normal(0.0005, 0.02, size=300)))
tsf = RollingTSF(window_size=20)(price)   # one-step-ahead value of the 20-bar trend line

fig = make_subplots(rows=2, cols=1, shared_xaxes=True,
                    row_heights=[0.55, 0.45], vertical_spacing=0.08)
fig.add_trace(go.Scatter(y=price, mode="lines", name="price"), row=1, col=1)
fig.add_trace(go.Scatter(y=tsf, mode="lines", name="TSF",
                         line=dict(color="red")), row=2, col=1)
fig.update_layout(title="Time-series forecast over 20 bars (RollingTSF)",
                  margin=dict(l=20, r=20, t=60, b=20),
                  legend=dict(orientation="h", yanchor="bottom", y=1.02, xanchor="right", x=1))
fig.update_yaxes(title_text="price", row=1, col=1)
fig.update_yaxes(title_text="forecast", row=2, col=1)
fig.show()