Why last beta is always positive? about piecewise_linear_fit_py HOT 3 OPEN

I did some simple test with 2-segment piecewise linear regression (unspecified breakpoints). I could output the breakpoints, which looked correct, and then I obtained yHat[0], (yHat[breakpointidx] - yHat[0]) / breakpointidx, (yHat[-1] - yHat[breakpointidx]) / (len(yHat) - breakpointidx), which should be expected to be very close to beta[0], beta[1], beta[2]. However, I could see the value of beta[-1] was different from (yHat[-1] - yHat[breakpointidx]) / (len(yHat) - breakpointidx) even by the sign.

from piecewise_linear_fit_py.

The source code appears all solid to me. I used fitfast but tested fit too -- both called fit_with_breaks, which altered beta via lstsq, seemingly a wrapper for scipy.linalg.lstsq.

I guess the issue is probably not a bug in the code, but in my interpretation. Any comments would be appreciated.

from piecewise_linear_fit_py.

beta[-1] can be negative! You just haven't had a curve with a steep enough slope yet at the end. The final slope needs to cancel out the accumulation of all previous slopes. Slopes on the right are the accumulation of all previous beta values (except the y-intercept aka. beta[0]).

See

import numpy as np
import matplotlib.pyplot as plt
import pwlf

# generate sin wave data
x = np.linspace(0, 11, num=100)
y = np.sin(x * np.pi / 2)
# add noise to the data
y = np.random.normal(0, 0.05, 100) + y

my_pwlf_1 = pwlf.PiecewiseLinFit(x, y, degree=1)

# fit the data for four line segments
res1 = my_pwlf_1.fitfast(5, pop=50)

# predict for the determined points
xHat = np.linspace(min(x), max(x), num=10000)
yHat1 = my_pwlf_1.predict(xHat)

print('beta', my_pwlf_1.beta)

# plot the results
plt.figure()
plt.plot(x, y, 'o', label='Data')
plt.plot(xHat, yHat1, '--', label='degree=1')
plt.legend()
plt.show()

beta [ 1.06781215 -0.5918096 1.87834207 -2.49662666 2.37475738 -2.42097531]

from piecewise_linear_fit_py.