Choosing and Evaluating Fits for Bond Yield Data

# Which curve is better to approximate bond yields (python)


# Which curve is better to approximate bond yields (python)












### I would like to approximate bond yields in python. But the question arose which curve describes this better?

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

x = [0.02, 0.22, 0.29, 0.38, 0.52, 0.55, 0.67, 0.68, 0.74, 0.83, 1.05, 1.06, 1.19, 1.26, 1.32, 1.37, 1.38, 1.46, 1.51, 1.61, 1.62, 1.66, 1.87, 1.93, 2.01, 2.09, 2.24, 2.26, 2.3, 2.33, 2.41, 2.44, 2.51, 2.53, 2.58, 2.64, 2.65, 2.76, 3.01, 3.17, 3.21, 3.24, 3.3, 3.42, 3.51, 3.67, 3.72, 3.74, 3.83, 3.84, 3.86, 3.95, 4.01, 4.02, 4.13, 4.28, 4.36, 4.4]
y = [3, 3.96, 4.21, 2.48, 4.77, 4.13, 4.74, 5.06, 4.73, 4.59, 4.79, 5.53, 6.14, 5.71, 5.96, 5.31, 5.38, 5.41, 4.79, 5.33, 5.86, 5.03, 5.35, 5.29, 7.41, 5.56, 5.48, 5.77, 5.52, 5.68, 5.76, 5.99, 5.61, 5.78, 5.79, 5.65, 5.57, 6.1, 5.87, 5.89, 5.75, 5.89, 6.1, 5.81, 6.05, 8.31, 5.84, 6.36, 5.21, 5.81, 7.88, 6.63, 6.39, 5.99, 5.86, 5.93, 6.29, 6.07]

a = np.polyfit(np.power(x,0.5), y, 1)
y1 = a[0]*np.power(x,0.5)+a[1]

b = np.polyfit(np.log(x), y, 1)
y2 = b[0]*np.log(x) + b[1]

c = np.polyfit(x, y, 2)
y3 = c[0] * np.power(x,2) + np.multiply(c[1], x) + c[2]

plt.plot(x, y, 'ro', lw = 3, color='black')
plt.plot(x, y1, 'g', lw = 3, color='red')
plt.plot(x, y2, 'g', lw = 3, color='green')
plt.plot(x, y3, 'g', lw = 3, color='blue')
plt.axis([0, 4.5, 2, 8])
plt.rcParams['figure.figsize'] = [10, 5]
```

The parabolic too goes down at the end (blue), the logarithmic goes too quickly to zero at the beginning (green), and the square root has a strange hump (red). Is there any other ways of more accurate approximation or is it that I'm already getting pretty good?

## Answer by oliversm (score 1, accepted)

https://quant.stackexchange.com/a/55685

## What makes a particular fit good?

Naturally for quantitative finance (and really for any quantitative science), while we can sometimes assess the performance of something by eye, this is not a reliable metric. Typically assessing something by eye is only acceptable when throwing away bad/awful fits. You should not use such qualitative methods for assessing or comparing models which do a reasonable job. As each of your proposed fits seems reasonable at a glance, this means you should not pursue qualitative methods any more.

Now that you want a quantitative method for assessing the performance of the various fits, there are a few different routes you could take. It is important to highlight that these are competing criteria, sometimes contradictory, and at this point it is important to know what your priorities are. (Ironically although I have said it's time for quantitative methods, choosing which method is more of an art than a science). To name a few you could do: Which has the least squared error (minimal $l^2$-error), $l^1$-error, $l^\infty$-error? Which model has the best AIC/BIC? Which performs better on a cross validation? Which is nicer analytically/mathematically for the desired purpose?

These were just some mathematical considerations, some practical ones might be which is simplest, which is quickest numerically, which is stable or sensitive to outliers, etc. Which can be done in real time, which are at risk of over fitting.

## Are all points created equal?

From a statistical or financial standpoint, it is also worth noting that not all points are created equal. By this I mean of course some are more important than others when it comes to fitting them correctly. Perhaps some of these are more liquid than others, perhaps some are higher volume, etc. This could lead to a weighted least squares, some specific loss function, etc.