Fitting models to weight loss data

Thu 12 February 2026

Function minimization, and analysis of diet results¶

Introduction¶

Halfway through last year (2025), I decided (at long last) to get serious about shedding weight.

I was eating healthy food, just too much of it. In order to keep motivated, I collected my daily weighin data, and produced a plot each day, predicting when my target weight would be achieved. I actually had two targets weights: 80 kilogram (kg) - a nice round number, and 78 kg (chosen after I got to 80 kg, and found my BMI still wasn't in the green zone).

I achieved my stretch goal on the morning on Christmas Day. Of course, my diet went out the window during the Christmas Day feast, but I am clawing back to 78 kg as I write these words

I found that the analysis of my diet's progress can be done in two different ways, each telling a different story, one more nuanced than the other.

Implementation¶

Imports¶

Import packages used

pathlib for file name management
pandas for DataFrame management
plotnine for plotting (my go-to package)
scipy for function minimization
statsmodels for fitting a linear model to the data

In [1]:

import pathlib

import numpy as np
import pandas as pd
import plotnine as p9
import scipy
import statsmodels.api as sm
from statsmodels.formula.api import ols

watermark provides reproducability information

In [2]:

%load_ext watermark

Data load¶

We read our data from an Excel spreadsheet, specifying:

the sheet in the spreadsheet we want
the columns to use
the initial rows to skip
the number of rows to read in
the names to use for the columns

In [3]:

path = pathlib.Path("diet_results.xlsx")
df = pd.read_excel(
    path,
    sheet_name="Data Entry",
    usecols=[3, 4],
    skiprows=2,
    nrows=141,
    names=["day", "weight"],
)

We examine the first and last few rows, and check the datatypes of the columns (all OK)

In [4]:

df.head()

Out[4]:

	day	weight
0	1	83.8
1	2	84.2
2	3	83.7
3	4	84.5
4	5	83.5

In [5]:

df.tail()

Out[5]:

	day	weight
136	173	78.6
137	174	78.5
138	175	78.3
139	176	78.2
140	177	78.0

In [6]:

df.dtypes

Out[6]:

day         int64
weight    float64
dtype: object

Data visualization¶

We use plotnine for our plotting.

We :

create an empty plot, specifying the default source of data for subsequent plotting operations
create a layer with points, with x, y values take from our source DataFrame
create a layer with one point, from the last row of the DataFrame, and show as a red point
use a black-and-white theme to miminize chart clutter
draw a horizontal line to show the final target weight
draw a vertical line to show when the diet stated (day 0)
add a plot subtitle, and caption

In [7]:

plot = (
    p9.ggplot(
        data=df,
    )
    + p9.geom_point(
        mapping=p9.aes(x="day", y="weight"),
    )
    + p9.geom_point(
        data=df.tail(1),
        mapping=p9.aes(x="day", y="weight"),
        color="red",
        size=3,
    )
    + p9.theme_bw()
    + p9.geom_hline(
        mapping=p9.aes(yintercept=78),
        color="gray",
    )
    + p9.geom_vline(
        mapping=p9.aes(xintercept=0),
        color="gray",
    )
    + p9.labs(subtitle="Diet analysis", caption="target weight achieved!")
)
plot

No description has been provided for this image

Plot enhancements¶

One of the nice features of plotnine is that you can add addition layers to the initial graphic.

In the next plot, we add a layer showing linear model (constant weight loss each day), with the 95% confidence area shaded grey. The model is matched the data by Ordinary Least Squares (ols)

In [8]:

plot = plot + p9.geom_smooth(mapping=p9.aes(x="day", y="weight"), method="ols", se=0.95)
plot

This tells an impressive story of constant improvment each day. A story of steady progress towards the target. To these skeptical people want statistical proof, we run the statsmodel ols routine, to get the p-values for each coefficient, and the overall R^2 value. We invert the weight loss per day to give "days to lose 1 kg": a more understandable concept than "grams lost each day".

In [9]:

res1 = ols("weight ~ day", data=df).fit()

# create a summary of the OLS fit to go on the plot
equation = (
    f'weight = {res1.params["Intercept"]:.1f} - day / {1/abs(res1.params["day"]):.1f}\n\n'
    + f"R^2 = {res1.rsquared:.2f}"
)

Show the statistical summary

In [10]:

plot = plot + p9.annotate(
    "text",
    x=100,
    y=84,
    label=equation,
    size=10,
    ha="left",
)
plot

A more nuanced model¶

I know that my motivation flagged in the run up to Christmas, and my dieting wasn't as strict as initially efforts.

So I decided to fit a more complex model. One with a steady linear decrease in weight, until a breakpoint was reached, and then no weight loss after that point.

I used the scipy.optimize.differential_evolution function.

The differential evolution method [1] is stochastic in nature. It does not use gradient methods to find the minimum, and can search large areas of candidate space, but often requires larger numbers of function evaluations than conventional gradient-based techniques.

I developed a function that took a set of parameters that characterised my model (y intercept, initial slope, and breakpoint after which slope=0), and computed the sum of squares of the deviation between model and actual data. You tell the scipy.optimize.differential_evolution function the range of each parameter, and it finds the parameter set that minimizes the sum-of-square-errors value (computed by the my_sq_error function.

Get the day and weight arrays into more conventional terms (x, y), as global arrays.

In [11]:

x = df["day"]
y = df["weight"]

The parameters set is (initial slope, y intercept , x breakpoint ). We define a function that takes this parameter set, and computes the sum of squared errors, model versus true data.

In [12]:

def my_sq_error(a):
    """
    my_sq_error: compute sum of square errors, model vs data

    Parameters: a - an array of model parameters

    We expect a to be an array 3 long, and we interpret the parameters as:
    a1 - initial slope (weight loss per day)
    a2 - initial weight
    a3 - day at which the daily weight loss was 0

    For each x value, we compute the model prediction, and square the error with
    the measured value.
    Finally, we sum all square error values

    We expect this function to be called by the function minimization utility
    """
    a1 = a[0]
    a2 = a[1]
    a3 = a[2]

    # value of estimated y at breakpoint x (= a3)
    a4 = a1 * a3 + a2

    sq_error = []
    for x_i, y_i in zip(x, y):
        y_est = a4 if x_i >= a3 else (a1 * x_i + a2)
        sq_error.append((y_est - y_i) ** 2)
    # end for
    return sum(sq_error)


# end my_sq_error

Specify the allowed range for each parameter (I have selected very generous ranges). This also tells the minimization utility how many parameters are in our model.

In [13]:

# input vector to D_E is [slope range, y intercept range, x breakpoint range]
slope_range = (-1.0, 1.0)  # says slope can be lossing a kg a day to gaining a kg a day!
yintercept_range = (50, 100)  # initial weight can vary from 50kg to 100kg
xbreakpoint_range = (
    100,
    200,
)  # the range of the day value after which no weight loss occurs
bounds = [slope_range, yintercept_range, xbreakpoint_range]

Run the minimization¶

In [14]:

result1 = scipy.optimize.differential_evolution(
    my_sq_error,
    bounds=bounds,
)

Get the new equation for weight loss

In [15]:

slope = result1["x"][0]
initial = result1["x"][1]
xbreakpoint = result1["x"][2]

In [16]:

equation = f"weight = {initial:.1f} - days / {(1/abs(slope)):.1f}"
equation

Out[16]:

'weight = 84.1 - days / 26.6'

Compute the predictions of our optimized model with final parameter set, and construct a new DataFrame

In [17]:

a1 = result1["x"][0]  # weight loss per day
a2 = result1["x"][1]  # initial weight
a3 = result1["x"][2]  # breakpoint day, after which no weight loss


a4 = a1 * a3 + a2  # predicted weight at breakpoint day

y_est = [a4 if x_i > a3 else (a1 * x_i + a2) for x_i in x]

df = pd.DataFrame({"x": x, "y": y, "y_est": y_est})

Visualize results¶

We:

create and empty plot, and specify our data source
create a layer with data points
create a layer with a vertical line at day=0
create a layer with a horizontal line at target weight
create a layer with a single point at the end of the weight loss line, larger, and colored tan
specify a black-and=white theme to minimize chart junk
create a layer with a line showing our model predictions
create a legend that explains the colors on the plot
label the x, y axis, and set the subtitle
show the model equation
resize the plot, now we have a legend taking up space to right

In [18]:

plot = (
    p9.ggplot(data=df)
    + p9.geom_point(mapping=p9.aes(x="x", y="y"))
    + p9.geom_vline(mapping=p9.aes(xintercept=0), color="gray")
    + p9.geom_hline(
        mapping=p9.aes(yintercept=78),
        color="gray",
    )
    + p9.geom_point(
        mapping=p9.aes(
            x=a3,
            y=a4,
            color='"tan"',
        ),
        size=5,
    )
    + p9.theme_bw()
    + p9.geom_line(
        mapping=p9.aes(x="x", y="y_est", color='"red"'),
    )
    + p9.scale_color_identity(
        name="Legend",
        guide="legend",
        breaks=["red", "tan"],
        labels=["Estimated slope", "Est. breakpoint"],
    )
    + p9.labs(
        x="day",
        y="weight",
        subtitle="Linear model with breakpoint",
    )
    + p9.annotate(
        "text",
        x=90,
        y=84,
        label=equation,
        size=10,
        ha="left",
    )
    + p9.theme(figure_size=(10, 6))
)
plot

Conclusions¶

Visually, the second model is more compelling. However, it would be a little galling (even if true) to say that my achieving my target was only due to random negative noise cutting in. If you look very carefully, the last week was a sprint to the finish line, with a consistent linear trend to get to 78kg before the Christmas feasting began!

Overall, I learnt that while it is easy to fit a linear model and say "all done" after you check the R^2 value, it is also moderately easy to use and fit more realistic models

Model checking¶

This section just checks that indeed the linear model with breakpoint has a smaller sum-of-errors-squared value than a pure linear model.

Define my sum-of-errors-squared function for the pure linear model

In [19]:

def my_sq_error_linear(a):
    """
    my_sq_error: compute sum of square errors, model vs data

    Parameters: a - an array of model parameters

    We expect a to be an array 2 long, and we interpret the parameters as:
    a1 - initial slope (weight loss per day)
    a2 - initial weight

    For each x value, we compute the model prediction, and square the error with
    the measured value.
    Finally, we sum all square error values

    We expect this function to be called by the function minimization utility
    """
    a1 = a[0]
    a2 = a[1]

    sq_error = []
    for x_i, y_i in zip(x, y):
        y_est = a1 * x_i + a2
        sq_error.append((y_est - y_i) ** 2)
    # end for
    return sum(sq_error)


# end my_sq_error_linear

Find the parameter set the minimizes this function

In [20]:

bounds = [
    slope_range,
    yintercept_range,
]
result2 = scipy.optimize.differential_evolution(
    my_sq_error_linear,
    bounds=bounds,
)

Display the values for each model: Linear Model with Breakpoint wins!

In [21]:

print(f"Sum of errors squared (Pure Linear Model)            = {result2['fun']:.4}")
print(f"Sum of errors squared (Linear Model with breakpoint) = {result1['fun']:.4}")

Sum of errors squared (Pure Linear Model)            = 20.87
Sum of errors squared (Linear Model with breakpoint) = 16.32

Finally, check that the Ordinary Least Square result from statsmodels (OLS) gives the same results as our function minimization of the pure linear model (FunMin): it does!

In [22]:

print(
    f" Intercept          : OLS={res1.params['Intercept']:.4}  vs FunMin = { result2['x'][1]:.4}\n"
    + f" Slope (days per kg): OLS={1/res1.params['day']:.4} vs FunMin = {1/result2['x'][0]:.4}"
)

 Intercept          : OLS=83.98  vs FunMin = 83.98
 Slope (days per kg): OLS=-28.51 vs FunMin = -28.51

In [ ]:

Reproducability information¶

In [23]:

%watermark

Last updated: 2026-02-12T13:52:43.708230+10:00

Python implementation: CPython
Python version       : 3.11.14
IPython version      : 9.10.0

Compiler    : MSC v.1929 64 bit (AMD64)
OS          : Windows
Release     : 10
Machine     : AMD64
Processor   : Intel64 Family 6 Model 170 Stepping 4, GenuineIntel
CPU cores   : 22
Architecture: 64bit

In [24]:

%watermark -h -iv -co

conda environment: fun_minim

Hostname: INSPIRON16

numpy      : 2.4.1
pandas     : 2.3.3
plotnine   : 0.15.0
scipy      : 1.16.3
statsmodels: 0.14.6

In [25]:

import contextlib

import ipynbname

with contextlib.suppress(FileNotFoundError):
    print(f"Notebook file name: {ipynbname.name()}")
# end with

Notebook file name: demo04