16 min read

Media Effect Estimation with Orbit's KTR Model

In this notebook we want to experiment to the new KTR model included in the new orbit’s release (1.1). In particular, we are interested in its applications to media effects estimation in the context of media mix modeling. This is one of the applications for the KTR model by the Uber’s team, see the corresponding paper Edwin, Ng, et al. “Bayesian Time Varying Coefficient Model with Applications to Marketing Mix Modeling”.

KTR stands for kernel-based Time-varying regression as its main feature is to fit time-variant regression coefficients in a bayesian forecasting model. For details on the KTR model, see the paper mentioned above and KTR documentation where there is a great sequence of tutorials. In a nutshell, given a time series \(y_t\) the KTR models its components as \(y_t = l_t + s_t + r_t + \varepsilon_t\), where \(l_t\) is the local-trend (level) , \(s_t\) is the seasonal component, \(r_t\) are the regression terms and \(\varepsilon_t\) is a stationary random error process. This is very similar to the Prophet model (we actually compare these two models bellow). One of the main differences is that the regression coefficients are time-varying. From the model documentation:

The core of KTR is the use of latent variables to define a smooth time varying representation of model coefficients. These smooth representations are Kernel Smooths.

These latent variables are somehow similar (or at least inspired) to the basis functions in a spline regression. Again, for details see the references above (which is strongly recommended!).

This notebook has two parts: - In Part I we generate the data ourselves. This allow us to have control on the data generation process in order to understand the strength and limitations of the KTR model. Moreover, we explore useful transformations of media cost data to model carryover and diminishing returns, see Jin, Yuxue, et al. “Bayesian methods for media mix modeling with carryover and shape effects.” (2017). - In Part II we use the KTR model to estimate the media effects. We also compare it with a baseline Prophet model.

Remark: In many applications measuring the effects via regression analysis is challenging because potential stochastic factors between the input and response variable. In our work Modelling stochastic time delay for regression analysis with Aaron Pickering we go deeper into this subject.

Prepare Notebook

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from orbit.models import KTR
from prophet import Prophet
from prophet.plot import add_changepoints_to_plot
from prophet.utilities import regressor_coefficients
from scipy.ndimage import gaussian_filter
from sklearn.base import BaseEstimator, TransformerMixin
import seaborn as sns

plt.style.use("bmh")
plt.rcParams["figure.figsize"] = [10, 6]
plt.rcParams["figure.dpi"] = 100
# orbit version
! pip list | grep -i "orbit"
orbit-ml    1.1.1.dev0

Part I: Generate Sample Data

We are going to generate a sample data set. The main idea is to generate a time series \(y\) which will represent target variable, lets say sales. This time series will consist of various components: a trend, a seasonal component, an external regressor and some gaussian noise. We will describe the components in more detail below. For now, let us tart by generating the data range: \(3\) years of weekly data.

min_date = pd.to_datetime("2018-04-01")
max_date = pd.to_datetime("2021-09-01")

data_df = pd.DataFrame(
    data={"date": pd.date_range(start=min_date, end=max_date, freq="W-MON")}
)

data_df = data_df.assign(
    year = lambda x: x["date"].dt.year,
    month = lambda x: x["date"].dt.month,
    dayofyear = lambda x: x["date"].dt.dayofyear,
)

n = data_df.shape[0]

Media Cost Features

  • Generate Media Cost Feature

Let us start by defining the external regressor: a media cost feature. This variable will play the role of sales driver for our target variable. In practice, the marketing manager will have control over this variable. In this example we use a uniform distribution to simulate it.

SEED = 42
np.random.seed(SEED)

# sample from a uniform distribution on [0, 1] and cut at the threshold (0.4)
z = np.random.uniform(low=0.0, high=1.0, size=n)
data_df["z"] = 10 * np.where( z > 0.4, z, z/4)

fig, ax = plt.subplots()
sns.lineplot(x="date", y="z", data=data_df, ax=ax)
ax.set(title="Raw Media Cost Data");
  • Adstock Transformation

Next we apply an adstock transformation to the media cost feature. This transformation aims to model the carry over effect of media spend. See Jin, Yuxue, et al. “Bayesian methods for media mix modeling with carryover and shape effects.” (2017) for more details.

Remark: In practice we would like to learn the adstock curve from the data. This can be done with PyMC as illustrated in the blog post Bayesian Marketing Mix Modeling in Python via PyMC3. This approach was also used in HelloFresh’s Media Mix Model, as presented in the following resources:

For this specific example we are going to assume a certain heuristic for the adstock curve, and hence it will not be part of the model. We are going to apply a geometric adstock transformation, which is defined by the following transformer:

class GeometricAdstockTransformer(BaseEstimator, TransformerMixin):

    def __init__(self, alpha: float = 0.0, l: int = 12):
        self.alpha = alpha
        self.l = l

    def fit(self, x, y=None):
        return self

    def transform(self, x):
        if isinstance(x, pd.DataFrame):
            x = x.to_numpy()
        cycles = [
            np.append(
                arr=np.zeros(shape=x.shape)[:i],
                values=x[: x.shape[0] - i],
                axis=0
            ) 
            for i in range(self.l)
        ]
        x_cycle = np.stack(cycles, axis=0)
        w = np.array([np.power(self.alpha, i) for i in range(self.l)])
        return np.tensordot(a=w, b=x_cycle, axes=1)

Next, lets apply the adstock transformation to the media cost feature.

# the effect carryover factor is 1/2 (usual assumption)
alpha = 1/2
# The carryover period is 12 weeks, which is also quite reasonable.
l = 12 

geometric_adstock_transformer = GeometricAdstockTransformer(alpha=alpha, l=l)

data_df["z_adstock"] = geometric_adstock_transformer.fit_transform(data_df["z"])

Let us plot the transformed data:

fig, axes = plt.subplots(
  nrows=2,
  ncols=1,
  sharex=True,
  sharey=False,
  layout="constrained"
)

features = ["z", "z_adstock"]

for i, (col, ax) in enumerate(zip(features, axes.flatten())):
    sns.lineplot(x="date", y=col, color=f"C{i}", label=col, data=data_df, ax=ax)
    ax.legend(loc="center left", bbox_to_anchor=(1, 0.5))

fig.suptitle("Media Cost Feature Transformations");
html
  • Saturation Transformation:

Now we apply a saturation transformation to the media cost feature, which should capture the non-linear effect of media in sales (otherwise the marketing manager would just spending more and more to increase sales, but experience and intuition shows the relation is non-linear and eventually saturates). There are many possible parametrizations for such a transformation, but we will use the following one:

\[ x\longmapsto \frac{1 - e^{-\mu x}}{1 + e^{-\mu x}} \]

Let us implement the corresponding transformer:

class LogisticSaturationTransformer(BaseEstimator, TransformerMixin):

    def __init__(self, mu: float = 0.5):
        self.mu = mu

    def fit(self, x, y=None):
        return self

    def transform(self, x):
        return (1 - np.exp(-self.mu * x)) / (1 + np.exp(-self.mu * x))

We can now apply the saturation transformation to the (adstock) media cost feature.

mu = 1.5e-1

logistic_saturation_transformer = LogisticSaturationTransformer(mu=mu)

data_df["z_adstock_saturated"] = logistic_saturation_transformer.fit_transform(data_df["z_adstock"])

We will not use z_adstock_saturated in our model and we would like to recover the saturation effect. Let us plot the saturation curve:

fig, ax = plt.subplots(figsize=(7, 6))

latex_function = r"$x\longmapsto \frac{1 - e^{-\mu x}}{1 + e^{-\mu x}}$"
sns.lineplot(
    x="z_adstock",
    y="z_adstock_saturated",
    color="C2",
    label=latex_function,
    data=data_df,
    ax=ax
)
ax.legend(loc="lower right", prop={"size": 18})
ax.set(title=f"Media Cost True Saturation Curve ($\mu$ = {mu})");

Note that the saturation curve is almost linear for values z_adstock < \(6\). After this point the first derivative starts to decrease.

Remark: It is clear that the adstock and saturation transformations do not commute (i.e. the order matters). Which one to apply first? Following Jin, Yuxue, et al. “Bayesian methods for media mix modeling with carryover and shape effects.” (2017) , section 2.3:

We could first apply the adstock transformation to the time series of media spend, and then apply the shape transformation. An alternative way would be to reverse the order. If media spend in each time period is relatively small compared to the cumulative spend across multiple time periods, the shape effect in each time period is less obvious compared to that of cumulative media spend. In this case, we would prefer to apply the shape transformation after the adstock transformation. However, if the media spend is heavily concentrated in some single time periods with an on-and-off pattern, the latter choice might be preferable.

Now we plot all the transformations:

fig, axes = plt.subplots(
  nrows=3,
  ncols=1,
  sharex=True,
  sharey=False,
  layout="constrained"
)

features = ["z", "z_adstock", "z_adstock_saturated"]

for i, (col, ax) in enumerate(zip(features, axes.flatten())):
    sns.lineplot(x="date", y=col, color=f"C{i}", label=col, data=data_df, ax=ax)
    ax.legend(loc='center left', bbox_to_anchor=(1, 0.5))

fig.suptitle("Media Cost Feature Transformations");
html
  • Diminishing Returns Over Time

Finally, we want to assume that the overall media efficiency decreases over time. We can do this by multiplying z_adstock_saturated with a decreasing function. We would expect to recover the decreasing effect with Orbit’s KTR model. Let us consider a rational function:

data_df["beta"] =  (np.arange(start=0.0, stop=1.0, step=1/n) + 1) ** (-1.8)

fig, ax = plt.subplots()
sns.lineplot(x="date", y="beta", color="C5", data=data_df, ax=ax)
ax.set(title="Media Cost Multiplicative Efficiency over Time", ylabel=None);

Next, we compute the product of beta and z_adstock_saturated:

data_df["z_effect"] = data_df["beta"] * data_df["z_adstock_saturated"]
fig, axes = plt.subplots(
    nrows=4,
    ncols=1,
    figsize=(12, 7),
    sharex=True,
    sharey=False,
    layout="constrained"
)

features = ["z", "z_adstock", "z_adstock_saturated", "z_effect"]

for i, (col, ax) in enumerate(zip(features, axes.flatten())):
    sns.lineplot(x="date", y=col, color=f"C{i}", label=col, data=data_df, ax=ax)
    ax.legend(loc='center left', bbox_to_anchor=(1, 0.5))

fig.suptitle("Media Cost Feature Transformations");
html

We can now plot the “true” saturation curves at yearly level based on the transformed data. We would like to recover the saturation effect. Note that we do not seed to expect a curve of the form \(x\longmapsto (1 - e^{-\mu x}) / (1 + e^{-\mu x})\) necessarily as we have a non-trivial time component (still, this parametrization could be a good approximation). For now, we simply use a lowess fit.

g = sns.lmplot(
    x="z_adstock",
    y="z_effect",
    hue="year",
    palette="viridis",
    lowess=True,
    scatter_kws={"edgecolor": "black", "alpha": 0.5},
    height=6,
    data=data_df
)
g.fig.suptitle("Media Cost Effect", y=1.02);

We can see from this plot that (1) the overall media efficiency is decaying over time and the (2) the saturation effect is present but not very strong.

Recall that for the model we will assume we have z_adstock and we would like to recover z_effect. Let us now plot the effect ratio defined by z_effect / z_adstock. We also smooth the curve (with a gaussian filter) to capture the overall trend.

data_df.eval(expr="effect_ratio = z_effect / z_adstock", inplace=True)

data_df["effect_ratio_smooth"] = gaussian_filter(input=data_df["effect_ratio"], sigma=8)

fig, ax = plt.subplots()
sns.lineplot(x="date", y="effect_ratio", color="black", label="effect ratio", data=data_df, ax=ax)
sns.lineplot(x="date", y="effect_ratio_smooth", color="C7", label="effect ratio smooth", data=data_df, ax=ax)
ax.set(title="Media Cost Effect Ratio");

Remark: Note that we applied the saturation and diminishing returns transformations independently. We could have done it differently by simply making \(\mu=\mu(t)\) time dependent in the LogisticSaturationTransformer. For example,

mu = np.linspace(start=0.05, stop=0.1, num=n)[::-1] ** (1.25)

This is the original formulation of Uber’s paper see section \(2\) in Edwin, Ng, et al. “Bayesian Time Varying Coefficient Model with Applications to Marketing Mix Modeling”. Some initial experiments show that ins this additive model structure, the saturation and diminishing returns transformations are well captured by the KTR model.

Trend and Seasonality

Next, we add some additive trend and seasonal components. For the trend component we simple use a rational power law:

data_df["trend"] = (np.linspace(start=0.0, stop=50, num=n) + 10)**(1/4) - 1 

For the seasonal component we use combinations of Fourier modes:

data_df["cs"] = - np.sin(2 * 2 * np.pi * data_df["dayofyear"] / 365.5) 
data_df["cc"] = np.cos(1 * 2 * np.pi * data_df["dayofyear"] / 365.5)
data_df["seasonality"] = 0.25 * (data_df["cs"] + data_df["cc"])

Let us plot the components:


fig, ax = plt.subplots()
sns.lineplot(x="date", y="trend", color="C4", label="trend", data=data_df, ax=ax)
sns.lineplot(x="date", y="seasonality", color="C6", label="seasonality", data=data_df, ax=ax)
ax.legend(loc="upper left")
ax.set(title="Trend & Seasonality Components", ylabel="");

Target Variable: Sales

As a final step in the data generation process let us put everything together. First, we can add an intercept and generate gaussian noise:

np.random.seed(SEED)

data_df["intercept"] = 3.0
data_df["trend_plus_intercept"] = data_df["trend"] + data_df["intercept"]

# the noise variance is increasing to make sure the resulting time series has constant variance
sigma_epsilon  = np.linspace(start=3e-2, stop=7e-2, num=n)
data_df["epsilon"] = np.random.normal(loc=0.0, scale=sigma_epsilon)

Now we are ready to build our additive time series:

data_df.eval(expr="y = intercept + trend + seasonality + z_effect + epsilon", inplace=True)

We can now plot resulting time series:

fig, ax = plt.subplots()
sns.lineplot(x="date", y="y", color="black", data=data_df, ax=ax)
ax.set(title="Sales (Target Variable)");

Let us compare the generated data with z_effect, which is the true effect of media on sales, and z_adstock, which is the the regressor we want to use in our regression model.

fig, axes = plt.subplots(nrows=3, ncols=1, sharex=True, sharey=False, layout="constrained")
sns.lineplot(x="date", y="y", color="black", data=data_df, ax=axes[0])
axes[0].set(title="Sales (Target Variable)")
sns.lineplot(x="date", y="z_effect", color="C3", data=data_df, ax=axes[1])
axes[1].set(title="Media Cost Effect on Sales")
sns.lineplot(x="date", y="z_adstock", color="C1", data=data_df, ax=axes[2])
axes[2].set(title="Media Cost Adstock Transform");

Part II: Modeling

In this second part of the notebook we are going to fit two time series model: (1) a Prophet model and (2) a KTR model. Even though we are mainly interested in the time varying coefficients provided by the KTR model, we start fit the Prophet model to get a baseline.

Prophet Model

The Prophet model is a good choice for time series with strong trend and seasonality patterns (see the great documentation and this blog post for an introduction). However, the regression coefficients do not vary over time. We use the default MAP estimation method to fit the model (not full MCMC).

  • Model Specification and Fit
# set model parameters
prophet_model = Prophet(yearly_seasonality=True)
prophet_model.add_regressor(name="z_adstock", mode="additive", standardize=False)
# fit model
prophet_model.fit(df=data_df.rename(columns={"date": "ds"}))
# generate predictions
prophet_pred_df = prophet_model.predict(df=data_df.rename(columns={"date": "ds"}))

First, let us see the model fit:

fig, ax = plt.subplots()
sns.lineplot(x="date", y="y", color="black", label="y", data=data_df, ax=ax)
fig = prophet_model.plot(fcst=prophet_pred_df, ax=ax)
a = add_changepoints_to_plot(ax=ax, m=prophet_model, fcst=prophet_pred_df)
ax.legend(loc="upper right")
ax.set(title="Prophet Model Fit");
  • Model Components

The fit looks overall quite good. Now let us visualize the model predicted and true components:

fig = prophet_model.plot_components(fcst=prophet_pred_df)

axes = fig.axes

sns.lineplot(
    x="ds",
    y="trend_plus_intercept",
    color="black",
    data=data_df.rename(columns={"date": "ds"}),
    label="true trend",
    ax=axes[0]
)
# seasonality plot is generated by plotting it in 2017 (source code) ¯\_(ツ)_/¯
sns.lineplot(
    x="ds_1",
    y="seasonality",
    color="black",
    data=(
        data_df
        .rename(columns={"date": "ds"})
        .query("year == 2019")
        .assign(ds_1 = lambda x: x["ds"]  - pd.DateOffset(years=2))
    ),
    label="true seasonality",
    ax=axes[1]
)
sns.lineplot(
    x="ds",
    y="z_effect",
    color="black",
    data=data_df.rename(columns={"date": "ds"}),
    label="z_effect",
    ax=axes[2]
)
fig.suptitle("Prophet Model Components", y=1.02);

The trend and seasonal components seem to be well captured by the model. On the other hand, the model does not capture the effect of media on sales, as expected.

  • Estimated Media Effect

We can extract the estimated beta coefficient for z_adstock and compare it with the true value.

prophet_regressor_df = regressor_coefficients(m=prophet_model)
beta_hat_prophet = prophet_regressor_df.query("regressor == 'z_adstock'")["coef"].to_numpy()[0]

fig, ax = plt.subplots()
sns.lineplot(
    x="date",
    y="effect_ratio",
    color="black",
    label="effect ratio",
    data=data_df,
    ax=ax
)
sns.lineplot(
    x="date",
    y="effect_ratio_smooth",
    color="C7",
    label="effect ratio smooth",
    data=data_df,
    ax=ax
)
ax.axhline(
    y=data_df["effect_ratio"].median(),
    color="black",
    linestyle="--",
    label=f"effect ratio median = {data_df['effect_ratio'].median(): 0.2f}"
)
ax.axhline(
    y=data_df["effect_ratio"].mean(),
    color="gray",
    linestyle="--",
    label=f"effect ratio mean = {data_df['effect_ratio'].mean(): 0.2f}"
)
ax.axhline(
    y=beta_hat_prophet,
    color="C1",
    linestyle="--",
    label="$\hat{\\beta}_{prophet}$ " + f"= {beta_hat_prophet: 0.3f}"
)
ax.legend(loc="upper right")
ax.set(title="Media Cost Effect Ratio");

In this example we see that the estimated beta coefficient for z_adstock in the Prophet model is close to the median of the true time-varying beta coefficient expected from the data generation process.

KTR Model

Now we fit the KTR model. We use the pyro-svi estimator which is a Pyro-based stochastic variational inference.

  • Model Specification and Fit
# define model parameters
date_col = "date"
response_col = "y"
regressor_col = ["z_adstock"]

ktr = KTR(
    response_col=response_col,
    date_col=date_col,
    regressor_col=regressor_col,
    regression_knot_distance=12,
    prediction_percentiles=[2.5, 97.5],
    seasonality=[52.25],
    regressor_sign=["+"],
    seed=42,
    estimator='pyro-svi',
)
# fit model
ktr.fit(df=data_df)

Let us generate predictions for the KTR model:

pred_df = ktr.predict(df=data_df, decompose=True)
fig, ax = plt.subplots()
sns.lineplot(x="date", y="y", color="black", label="y", data=data_df, ax=ax)
ax.fill_between(
    x=pred_df["date"],
    y1=pred_df["prediction_2.5"],
    y2=pred_df["prediction_97.5"],
    color="C0",
    alpha=0.15,
    label="prediction interval (95%)"
)
sns.lineplot(
  x="date",
  y="prediction",
  color="C0",
  label="prediction",
  data=pred_df,
  ax=ax
)
sns.lineplot(
    x="date",
    y="trend",
    alpha=0.8,
    color="red",
    label="fitted trend",
    data=pred_df,
    ax=ax
)
sns.scatterplot(
    x="date",
    y="lev_knot",
    data=ktr.get_level_knots(),
    s=50,
    color="red",
    label="level knots",
    ax=ax
)
for row in ktr.get_level_knots().itertuples():
    ax.axvline(x=row.date, color="red", alpha=0.3, linestyle="--")

ax.legend(loc="center left", bbox_to_anchor=(1, 0.5))
ax.set(title="KTR Model Fit");
html

We can get the complete posterior samples as follows:

for k, samples in ktr.get_posterior_samples().items():
    print(k, samples.shape)
lev_knot (100, 11)
lev (100, 179)
yhat (100, 179)
obs_scale (100,)
coef_knot (100, 16)
coef_init_knot (100,)
coef (100, 179)
  • Model Components

Next, we can plot the fitted trend ans seasonal components:

fig, ax = plt.subplots(
  nrows=2,
  ncols=1,
  sharex=True,
  sharey=False,
  layout="constrained"
)

sns.lineplot(x="date",
    y="seasonality",
    color="C4",
    label="s",
    data=data_df,
    ax=ax[0]
)
ax[0].fill_between(
    x=pred_df["date"],
    y1=pred_df["seasonality_52.25_2.5"],
    y2=pred_df["seasonality_52.25_97.5"],
    color="C0",
    alpha=0.15,
    label="prediction interval (95%)"
)
sns.lineplot(
    x="date",
    y="seasonality_52.25",
    alpha=0.8,
    color="C10",
    label="fitted seasonality",
    data=pred_df,
    ax=ax[0]
)
ax[0].legend(loc="upper left")
ax[0].set(title="Seasonality Fit", ylabel=None)

sns.lineplot(
    x=data_df["date"],
    y=data_df["trend"] + data_df["intercept"],
    color="C4",
    label="trend + intercept",
    ax=ax[1]
)
ax[1].fill_between(
    x=pred_df["date"],
    y1=pred_df["trend_2.5"],
    y2=pred_df["trend_97.5"],
    color="C0",
    alpha=0.15,
    label="prediction interval (95%)"
)
sns.lineplot(
    x="date",
    y="trend",
    alpha=0.8,
    color="C0",
    label="fitted trend",
    data=pred_df,
    ax=ax[1]
)
sns.scatterplot(
    x="date",
    y="lev_knot",
    data=ktr.get_level_knots(),
    s=50,
    label="level knots",
    ax=ax[1]
)
ax[1].legend(loc="upper left")
ax[1].set(title="Trend Fit", ylabel=None)
fig. suptitle("KTR Model Components", y=1.04);

The fit looks a bit better than the prophet model (specially the trend component).

  • Estimated Media Effect

Now let us extract the beta coefficient for z_adstock:

coef_mid, coef_lower, coef_upper  = ktr.get_regression_coefs(include_ci=True)

From the estimated beta coefficient we can compute the estimated media effect on sales:

for df in coef_mid, coef_lower, coef_upper:
    df["z_effect_pred"] = df["z_adstock"] * data_df["z_adstock"]
    df["year"] = df["date"].dt.year # this is a feature used later for viz purposes

We can now compare it with the true effect ratio:

fig, ax = plt.subplots(
    nrows=3,
    ncols=1,
    figsize=(11, 8),
    sharex=True,
    sharey=False,
    layout="constrained"
)

sns.lineplot(
    x="date",
    y="z_adstock",
    color="C1",
    label="z_effect",
    data=data_df,
    ax=ax[0]
)
ax[0].set(title="Media Cost Adstock Transform")

sns.lineplot(
    x="date",
    y="effect_ratio",
    color="black",
    data=data_df,
    ax=ax[1]
)
ax[1].fill_between(
    x=coef_mid["date"],
    y1=coef_lower["z_adstock"],
    y2=coef_upper["z_adstock"],
    color="C9",
    label="95% credible interval",
    alpha=0.15
)
sns.lineplot(
    x="date",
    y="z_adstock",
    color="C9",
    label="$\hat{\\beta}_{z}$",
    data=coef_mid,
    ax=ax[1]
)
sns.scatterplot(
    x="date",
    y="z_adstock",
    data=ktr.get_regression_coef_knots(),
    s=50,
    label="regression knots",
    ax=ax[1]
)
ax[1].set(title="Effect Ratio")

ax[2].fill_between(
    x=coef_mid["date"],
    y1=coef_lower["z_effect_pred"],
    y2=coef_upper["z_effect_pred"],
    color="C9",
    label="95% credible interval",
    alpha=0.15
)
sns.lineplot(x="date",
    y="z_effect_pred",
    color="C9",
    label="z_effect pred",
    data=coef_mid,
    ax=ax[2]
)
sns.lineplot(x="date",
    y="z_effect",
    color="C7",
    label="z_effect",
    data=data_df,
    ax=ax[2]
)
ax[2].set(title="Media Effect Cost Ratio ($\\beta$-coef) KTR Model");
html

The KTR model captures the overall diminishing returns effect over time. Let us now zoom in and compare the effect estimation of the Prophet and KTR models:

fig, ax = plt.subplots()
sns.lineplot(
    x="date",
    y="effect_ratio",
    color="black",
    label="effect ratio",
    data=data_df,
    ax=ax
)
sns.lineplot(
    x="date",
    y="effect_ratio_smooth",
    color="C7",
    label="effect ratio smooth",
    data=data_df,
    ax=ax
)
ax.axhline(
    y=data_df["effect_ratio"].median(),
    color="black",
    linestyle="--",
    label=f"effect ratio median = {data_df['effect_ratio'].median(): 0.2f}"
)
ax.axhline(
    y=data_df["effect_ratio"].mean(),
    color="gray",
    linestyle="--",
    label=f"effect ratio mean = {data_df['effect_ratio'].mean(): 0.2f}"
)
ax.axhline(
    y=beta_hat_prophet,
    color="C1",
    linestyle="--",
    label="$\hat{\\beta}_{prophet}$ " + f"= {beta_hat_prophet: 0.3f}"
)
sns.lineplot(
    x="date",
    y="z_adstock",
    color="C0",
    label="$\hat{\\beta}_{KTR}$",
    data=coef_mid,
    ax=ax
)
ax.fill_between(
    x=coef_mid["date"],
    y1=coef_lower["z_adstock"],
    y2=coef_upper["z_adstock"],
    color="C0",
    label="$\hat{\\beta}_{KTR}$ 95% credible interval",
    alpha=0.15
)
ax.legend(loc="upper right")
ax.set(title="Media Cost Effect Ratio");
html

It is clear that the KTR model flexibility for time-varying coefficients outperforms the Prophet model base model. Nevertheless, the estimated effect by the KTR model is a bit lower than the true effect.

  • Saturation Effect

Finally, let us see how much we can recover the saturation effect:

fig, ax = plt.subplots(figsize=(9, 8))

markers, caps, bars = ax.errorbar(
    x=data_df["z_adstock"], 
    y=coef_mid["z_effect_pred"], 
    yerr=(coef_upper["z_effect_pred"] - coef_lower["z_effect_pred"]) / 2,
    fmt='o',
    ms=5,
    capsize=5,
    label="estimated effect",
)
[bar.set_alpha(0.3) for bar in bars]


latex_function = r"$x\longmapsto \frac{1 - e^{-\mu z}}{1 + e^{-\mu z}}$"
sns.lineplot(
    x="z_adstock",
    y="z_adstock_saturated",
    color="C2",
    label="upper bound " + latex_function,
    linestyle="--",
    data=data_df,
    ax=ax
)
sns.scatterplot(
    x="z_adstock",
    y="z_effect",
    color="C3",
    label="true effect",
    data=data_df,
    ax=ax
)
ax.legend(loc='center left', bbox_to_anchor=(1, 0.5))
ax.set(title="Media Cost Effect Estimation", ylabel="effect");
html

The dashed purple curve denotes the true saturation curve from the data generation process. Recall that the true effect was assumed to be product of the z_adstock_saturated by a decreasing function \(\beta = \beta(t) \subset [0, 1]\). These true estimated effects are the green points, which are of course below the dashed purple curve since the function \(\beta(t)\) is in particular less or equal than one. The blue points (and the credible intervals) denote the KTR estimated media effect on sales. From this plot we can not easily infer the saturation effect. The reason is that the time component is not explicitly represented in this plot. Let us split it by year:

fig, axes = plt.subplots(
    nrows=2,
    ncols=2,
    figsize=(10, 9), 
    sharex=True,
    sharey=True,
    layout="constrained"
)

axes = axes.flatten()

for i, year in enumerate(data_df["year"].sort_values().unique()):
    ax = axes[i]
    mask = f"year == {year}"
    
    yerr = coef_upper.query(mask)["z_effect_pred"] - coef_lower.query(mask)["z_effect_pred"]
    
    markers, caps, bars = ax.errorbar(
        x=data_df.query(mask)["z_adstock"], 
        y=coef_mid.query(mask)["z_effect_pred"], 
        yerr=yerr/2,
        fmt='o',
        ms=0,
        capsize=5,
        label="estimated effect",
    )
    [bar.set_alpha(0.3) for bar in bars]

    sns.regplot(
        x=data_df.query(mask)["z_adstock"], 
        y=coef_mid.query(mask)["z_effect_pred"], 
        lowess=True,
        color="C0",
        line_kws={"alpha": 0.5},
        ax=ax
    )

    sns.regplot(
        x="z_adstock",
        y="z_effect",
        lowess=True,
        color="C3",
        label="true effect",
        data=data_df.query(mask),
        scatter_kws={"alpha": 0.3},
        line_kws={"alpha": 0.3},
        ax=ax
    )
    if i == 0:
        ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.2), ncol=3)
    else:
        ax.legend().remove()
    ax.set(title=f"{year}")
    
fig.suptitle("Media Cost Effect Estimation KTR Model", y=1.05);
html

In this plot we also see that the KTR is underestimating the media effect on sales most of the time. For both, the true and estimated effects we are using a lowess interpolation. In this example, the true saturation effect is by itself not very strong. The KTR model does show a non-linear relation as seen by the lowess fit. The year \(2020\) shows the closes fit.