Exploring TensorFlow Probability STS Forecasting

Smoothing

We begin the analysis by studying and visualize various level of smoothness using a gaussian_filter. You can think of it as a weighted centered moving average based on a normal distribution. In particular, the level of smoothness is controlled by the parameter sigma which represents the standard deviation.

from scipy.ndimage import gaussian_filter

df_smooth = df_train \
    .assign(y_smooth_1 = lambda x: gaussian_filter(input=x['y'], sigma=3.5)) \
    .assign(y_smooth_2 = lambda x: gaussian_filter(input=x['y'], sigma=15)) \

fig, ax = plt.subplots()
sns.lineplot(x='date', y='y', label='y_train', data=df_smooth, alpha=0.5, ax=ax)
sns.lineplot(x='date', y='y_smooth_1', label='y_train_smooth monthly', data=df_smooth, ax=ax)
sns.lineplot(x='date', y='y_smooth_2', label='y_train_smooth yearly', data=df_smooth, ax=ax)
ax.legend(loc='upper left')
ax.set(title='Dependent Variable - Smoothing', ylabel='y');

From this plot we see two clear seasonalities: yearly and monthly. The variable y_train_smooth yearly also includes a positive trend component, which does not look linear.

To understand how to model the monthly seasonality we plot the mean of y over the day of the month.

fig, ax = plt.subplots()
df_train.groupby('day_of_month').agg({'y': np.mean}).plot(ax=ax)
ax.set(title='Day of the month mean');

This shows a good first approximation is to model this component via a cyclic variable of the form \(z\longmapsto \sin(z)\).

Remove yearly and monthly seasonality

Let us remove these two seasonal components:

# Remove yearly seasonality.
y_no_year_season = df_smooth['y'] - df_smooth['y_smooth_2']
# Remove monthly seasonality.
y_no_year_month_season = y_no_year_season \
    - gaussian_filter(input=y_no_year_season, sigma=3.5)

# Plot components.
fig, ax = plt.subplots(3, 1, figsize=(12, 16))
sns.lineplot(x='date', y='y', label='y_train', data=df_smooth, alpha=0.5, ax=ax[0])
sns.lineplot(x='date', y='y_smooth_1', label='y_train_smooth monthly', data=df_smooth, ax=ax[0])
sns.lineplot(x='date', y='y_smooth_2', label='y_train_smooth yearly', data=df_smooth, ax=ax[0])
ax[0].set(title='y')

ax[1].plot(df_smooth['date'], y_no_year_season, c=sns_c[2])
ax[1].axhline(y_no_year_season.mean(), color=sns_c[3], linestyle='--', label='mean')
ax[1].legend()
ax[1].set(title='y no year seasonality')

ax[2].plot(df_smooth['date'], y_no_year_month_season, c=sns_c[1])
ax[2].axhline(y_no_year_month_season.mean(), color=sns_c[3], linestyle='--', label='mean')
ax[2].legend()
ax[2].set(title='y no year & month seasonality');

AC and PAC

Let us now compute the autocorrelation and partial-autocorrelation of the remainder component.

from statsmodels.graphics.tsaplots import plot_acf, plot_pacf

fig, ax = plt.subplots(2, 1)
plot_acf(x=y_no_year_month_season, ax=ax[0])
plot_pacf(x=y_no_year_month_season, ax=ax[1]);

From the autocorrelation plot we observe a clear seasonality component at lag 7. Moreover, looking into the shape of the plot we can also see that the encoding is cyclical.

Correlations

We want to see whether there is a linear effect of x on y. A first good indication can be obtained by looking into correlations. Naively, one would simply compute:

np.corrcoef(df_train['y'], df_train['x'])[0, 1]

0.3773272976778308

However, this correlation does not reflect the potential relation of x on y since we have a clear positive trend and seasonality components. A more meaningful indication is the correlation:

np.corrcoef(y_no_year_month_season, df_train['x'])[0, 1]

0.6714258581031259

Note that we still have not removed the 7-day (weekly) seasonality. This of course has an effect on this correlation. To see this let us consider the scatter plot:

fig, ax = plt.subplots()
sns.scatterplot(
    x=y_no_year_month_season, 
    y=df_train['x'], 
    hue=df['day_of_week'],
    palette='viridis',
    ax=ax, 
)

ax.set(
    title='y_no_year_month_season vs x', 
    xlabel='x', 
    ylabel='y_no_year_month_season'
);

We indeed see we have clear clusters on each level of x corresponding to each day of the week.

One way of removing this seasonality is via a linear regression model. We use a one-hot encoding for the day of the week variable.

Remark: As we have seen before a cyclical encoding might be a better choice. Nevertheless, given how the data was generated, we want to test the one-hot encoding approach.

# Prepare model data frame.
dow_df = df_train[['day_of_week']].copy()
dow_df['y_no_year_month_season'] = y_no_year_month_season
# One-hot encoding of the day of the week.
dow_dummies = pd.get_dummies(dow_df['day_of_week'], drop_first=True)
dow_dummies.columns = ['d' + str(i) for i in dow_dummies.columns]
dow_df = pd.concat([dow_df, dow_dummies], axis=1)
dow_df.head()

Next, we fit the linear model.

import statsmodels.formula.api as smf

dow_mod = smf.ols(
    formula='y_no_year_month_season ~  d1 + d2 + d3 + d4 + d5 + d6', 
    data=dow_df
)

dow_res = dow_mod.fit()

print(dow_res.summary())

                        OLS Regression Results                              
==================================================================================
Dep. Variable:     y_no_year_month_season   R-squared:                       0.350
Model:                                OLS   Adj. R-squared:                  0.347
Method:                     Least Squares   F-statistic:                     114.2
Date:                    Tue, 11 Feb 2020   Prob (F-statistic):          2.64e-115
Time:                            11:32:43   Log-Likelihood:                -2214.7
No. Observations:                    1277   AIC:                             4443.
Df Residuals:                        1270   BIC:                             4479.
Df Model:                               6                                         
Covariance Type:                nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
Intercept      1.4182      0.102     13.958      0.000       1.219       1.618
d1            -0.6520      0.144     -4.538      0.000      -0.934      -0.370
d2            -1.7017      0.144    -11.826      0.000      -1.984      -1.419
d3            -2.7558      0.144    -19.151      0.000      -3.038      -2.473
d4            -2.6323      0.144    -18.293      0.000      -2.915      -2.350
d5            -1.7910      0.144    -12.446      0.000      -2.073      -1.509
d6            -0.4125      0.144     -2.871      0.004      -0.694      -0.131
==============================================================================
Omnibus:                      665.602   Durbin-Watson:                   2.277
Prob(Omnibus):                  0.000   Jarque-Bera (JB):             3977.387
Skew:                           2.431   Prob(JB):                         0.00
Kurtosis:                      10.150   Cond. No.                         7.86
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

We now compute the correlation of the regressor x with the model residuals:

np.corrcoef(dow_res.resid, df_train['x'])[0, 1]

0.8612502137457188

The correlation increased significantly.

Regressor Effect

A common question in time series analysis is to estimate the effect (ROI) of a specific external regressor (e.g. what is the ROI of media spend on sales?) For this specific case, given the correlation above, we can estimate this for x running a linear model on the residuals of the dow_model.

x_df = df_train[['x']].copy()
x_df['dow_model_resid'] = dow_res.resid

x_mod = smf.ols(formula='dow_model_resid ~ x', data=x_df)

x_res = x_mod.fit()

print(x_res.summary())

                      OLS Regression Results                            
==============================================================================
Dep. Variable:        dow_model_resid   R-squared:                       0.742
Model:                            OLS   Adj. R-squared:                  0.742
Method:                 Least Squares   F-statistic:                     3662.
Date:                Tue, 11 Feb 2020   Prob (F-statistic):               0.00
Time:                        11:32:44   Log-Likelihood:                -1350.3
No. Observations:                1277   AIC:                             2705.
Df Residuals:                    1275   BIC:                             2715.
Df Model:                           1                                         
Covariance Type:            nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
Intercept     -0.2907      0.020    -14.467      0.000      -0.330      -0.251
x              5.2360      0.087     60.515      0.000       5.066       5.406
==============================================================================
Omnibus:                       31.128   Durbin-Watson:                   1.226
Prob(Omnibus):                  0.000   Jarque-Bera (JB):               34.638
Skew:                          -0.340   Prob(JB):                     3.01e-08
Kurtosis:                       3.435   Cond. No.                         4.45
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

The estimates effect is then:

x_res.params['x']

5.236046936179519

Recall that the true effect is 5.0.

Periodogram

There is another way to extract seasonality patterns: using a periodogram to estimate the spectral density of a signal.

from scipy import signal

f, Pxx_den = signal.periodogram(
    x=df_train['y'], 
    detrend='linear', 
    nfft=int(7e2)
)

fig, ax = plt.subplots()
sns.lineplot(x=f, y=Pxx_den, ax=ax)
ax.set(title='Power Spectral Density');

Each of these peaks represent the seasonal components. To compute the associated frequency wee need to compute their multiplicative inverse.

# Sort to get the peak values.
sort_freq_index = np.argsort(a=Pxx_den)[::-1]

periodogram_df = pd.DataFrame(
    {'sort_freq': f[sort_freq_index], 'Pxx_den': Pxx_den[sort_freq_index]}
)

periodogram_df.assign(days = lambda x: 1/x['sort_freq']).head(5)

The first three values correspond to days = 30, 350 and 7. Which correspond to monthly, yearly and weekly seasonality.

Time Series Decomposition

statsmodels has an inbuilt decomposition function using moving averages. Let us use it to estimate the effect of the external regressor (which we already know it should be the residual component).

from statsmodels.tsa.seasonal import seasonal_decompose

decomposition_obj = seasonal_decompose(
    x=df_train[['date', 'y']].set_index('date'), 
    model='additive'
)

fig, ax = plt.subplots(4, 1, figsize=(12, 12))

decomposition_obj.observed.plot(ax=ax[0])
ax[0].set(title='observed')
decomposition_obj.trend.plot(ax=ax[1])
ax[1].set(title='trend')
decomposition_obj.seasonal.plot(ax=ax[2])
ax[2].set(title='seasonal')
decomposition_obj.resid.plot(ax=ax[3])
ax[3].set(title='residual')
plt.tight_layout()

Note that the trend component also includes the monthly and yearly seasonality. This might not always be desired. You can of course use the smoothing technique presented above or keep decomposing the resulting trend by specifying the period parameter in the seasonal_decompose function.

• Correlation

# We remove the first and last 3 entries as they are np.nan 
# coming from the moving average method. 
np.corrcoef(decomposition_obj.resid[3:-3], df_train['x'][3:-3])[0, 1]

0.8616875939727174

• Effect Estimation

x_df['decomposition_resid'] = decomposition_obj.resid.values

x_mod2 = smf.ols(formula='decomposition_resid ~ x', data=x_df[3:-3])

x_res2 = x_mod2.fit()

print(x_res2.summary())

                       OLS Regression Results                            
===============================================================================
Dep. Variable:     decomposition_resid   R-squared:                       0.743
Model:                             OLS   Adj. R-squared:                  0.742
Method:                  Least Squares   F-statistic:                     3659.
Date:                 Tue, 11 Feb 2020   Prob (F-statistic):               0.00
Time:                         11:32:47   Log-Likelihood:                -1311.4
No. Observations:                 1271   AIC:                             2627.
Df Residuals:                     1269   BIC:                             2637.
Df Model:                            1                                         
Covariance Type:             nonrobust                                         
==============================================================================
                 coef    std err          t      P>|t|      [0.025      0.975]
------------------------------------------------------------------------------
Intercept     -0.2828      0.020    -14.404      0.000      -0.321      -0.244
x              5.1026      0.084     60.492      0.000       4.937       5.268
==============================================================================
Omnibus:                       73.490   Durbin-Watson:                   1.441
Prob(Omnibus):                  0.000   Jarque-Bera (JB):               88.742
Skew:                          -0.568   Prob(JB):                     5.37e-20
Kurtosis:                       3.620   Cond. No.                         4.44
==============================================================================

Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

We get similar results as above.

Remarks:

• In real life applications the effects of external regressors might include lags. A correlation analysis with lags needs to be done before modeling.
• Also, linear relations might serve as a good first approximation. Nevertheless, going into non-linear relations might improve the model significantly. Two concrete methods which appear often is to model via adstock effect and generalized additive models.

Variational Inference (Short Intro)

We follow the strategy of the TensorFlow Probability use cases and fit the model using variational inference. I find the article Variational Inference: A Review for Statisticians very good for an introduction to the subject. The main idea of variational inference is to use optimization methods to minimize the Kullback-Leibler divergence in order to approximate a conditional density within a family of specified parameter densities \(\mathfrak{D}\). Specifically, assume we are given observed variables \(x\) (i.e. data) and let \(z\) be a set of latent variables with joint density \(p(x, z)\). We are interested in computing the conditional density

\[ p(z|x) = \frac{p(x, z)}{p(x)}, \quad\text{where} \quad p(x)=\int p(x, z) dz. \]

This last integral (known as the evidence) is in general very hard to compute. The main idea of variational inference is to solve the optimization problem

\[ q^{*}(z)=\min_{q \in \mathfrak{D}} \text{KL}(q(z)||p(z|x)), \]

where

\[ \begin{align} \text{KL}(q(z)||p(z|x)) =& \text{E}[\log(q(z))] - \text{E}[\log(p(z|x))] \\ =& \text{E}[\log(q(z))] - \text{E}[\log(p(z,x))] - \text{E}[\log(p(x))] \\ =& \text{E}[\log(q(z))] - \text{E}[\log(p(z,x))] - \log(p(x) \end{align} \]

is the Kullback-Leibler (KL) divergence (the expected values are taken with respect to \(q(z)\)). Note that this quantity contains a term \(\text{E}[\log(p(x))]\), which is hard to compute. Because we cannot compute the KL, we optimize an alternative objective that is equivalent to the KL up to an added constant ²,

\[ \text{ELBO}(q) = \text{E}[\log(p(z, x))] - \text{E}[\log(q(z))] \]

ELBO stands for evidence lower bound.

Remark: Note that

\[ \text{ELBO}(q)= -\text{KL}(q(z)||p(z|x)) + \log(p(x)) \]

Moreover, it is easy to see that

\[ \text{ELBO}(q)= \text{E}[\log(p(x|z)]-\text{KL}(q(z)||p(x)) \]

The first term is the expected likelihood and the second term is the negative divergence between the variational density and the prior.

For more details and enlightening comments, please refer to the article mentioned above (from which this short introduction was taken from).

Variational Inference in TensorFlow Probability

First we build the variational surrogate posteriors, which consist of independent normal distributions with two trainable hyper-parameters loc and scale for each parameter in the model. That is, \(\mathfrak{D}\) above consists of normal distributions.

variational_posteriors = tfp.sts.build_factored_surrogate_posterior(
  model=toy_model, 
  seed=42
)

Let us sample from the prior distributions and plot the corresponding densities:

q_prior_samples = variational_posteriors.sample(1000)

num_parameters = len(toy_model.parameters)

fig, ax = plt.subplots(num_parameters, 1, figsize=(12, 21))

for i, param in enumerate(toy_model.parameters):
    
    param_mean = np.mean(q_prior_samples[param.name], axis=0)
    param_median = np.median(q_prior_samples[param.name], axis=0)
    param_std = np.std(q_prior_samples[param.name], axis=0)
    
    sns.distplot(a=q_prior_samples[param.name], rug=True, ax=ax[i])
    ax[i].set(title=param.name)
    ax[i].axvline(x= param_median, color=sns_c[1], linestyle='--', label='median')
    ax[i].axvline(x= param_mean, color=sns_c[2], linestyle='--', label='mean')
    ax[i].axvline(x= param_mean + param_std, color=sns_c[3], linestyle='--', label=r'$\mu \pm 2\sigma$')
    ax[i].axvline(x= param_mean - param_std, color=sns_c[3], linestyle='--')
    ax[i].legend()
    
plt.tight_layout(pad=3.0)
plt.suptitle('PRIOR DISTRIBUTIONS', y=0.99);

Next we run the optimization procedure.

num_variational_steps = int(200)

# Set optimizer.
optimizer = tf.optimizers.Adam(learning_rate=0.1)

# Using fit_surrogate_posterior to build and optimize 
# the variational loss function.
@tf.function(experimental_compile=True)
def train():
    # Build the joint density. 
    target_log_prob_fn = toy_model.joint_log_prob(
        observed_time_series=df_train['y']
    )
    
    elbo_loss_curve = tfp.vi.fit_surrogate_posterior(
        target_log_prob_fn=target_log_prob_fn,
        surrogate_posterior=variational_posteriors,
        optimizer=optimizer,
        num_steps=num_variational_steps,
        seed=42
    )
    
    return elbo_loss_curve

# Run optimization.
elbo_loss_curve = train()

Let us plot the elbo_loss_curve.

fig, ax = plt.subplots()
ax.plot(elbo_loss_curve, marker='.')
ax.set(title='ELBO Loss Curve', xlabel='iteration');

We see that a minimum has been reached.

Now we sample from the variational posteriors obtained:

q_samples = variational_posteriors.sample(1000)

Let us plot the parameter posterior distributions:

num_parameters = len(toy_model.parameters)

fig, ax = plt.subplots(num_parameters, 1, figsize=(12, 21))

for i, param in enumerate(toy_model.parameters):
    
    param_mean = np.mean(q_samples[param.name], axis=0)
    param_median = np.median(q_samples[param.name], axis=0)
    param_std = np.std(q_samples[param.name], axis=0)
    
    sns.distplot(a=q_samples[param.name], rug=True, ax=ax[i])
    ax[i].set(title=param.name)
    ax[i].axvline(x= param_median, color=sns_c[1], linestyle='--', label='median')
    ax[i].axvline(x= param_mean, color=sns_c[2], linestyle='--', label='mean')
    ax[i].axvline(x= param_mean + param_std, color=sns_c[3], linestyle='--', label=r'$\mu \pm 2\sigma$')
    ax[i].axvline(x= param_mean - param_std, color=sns_c[3], linestyle='--')
    ax[i].legend()
    
plt.tight_layout(pad=3.0)
plt.suptitle('POSTERIOR DISTRIBUTIONS', y=0.99);

# Get mean and std for each parameter.
print('Inferred parameters:')
for param in toy_model.parameters:
    print('{}: {} +- {}'.format(
        param.name,
        np.mean(q_samples[param.name], axis=0),
        np.std(q_samples[param.name], axis=0))
    )

    Inferred parameters:
    observation_noise_scale: 0.5021326541900635 +- 0.007594980299472809
    local_linear_trend/_level_scale: 0.009583557024598122 +- 0.006971438881009817
    local_linear_trend/_slope_scale: 0.0018680891953408718 +- 0.0003248010470997542
    month_of_year/_drift_scale: 0.024883059784770012 +- 0.08357027918100357
    day_of_week/_drift_scale: 0.018588630482554436 +- 0.004957796074450016
    day_of_month/_weights: [1.9918456] +- [0.02078216]
    x_var/_weights: [5.9605927] +- [0.07460035]

Remark: Observe that the estimated effect of the regressor x is around 5.96.