Code#

You can find the Matlab (R2024b) files in this GitHub directory (still in beta version).

Below is documentation for the files. However, we suggest first reading the example in home, which is an easier way to get a general idea of how the code works.

Range of data accepted:#

Below list the current settings on the data dimensions accepted by the main file nne_estimate.m. (Note: current settings may seem restrictive, but We can change them in the future if there are requests. Please feel free to contact us.)

  • Sample size: \(n\) ≥ 1000 consumers (or search sessions). Number of options: 15 ≤ \(J\) ≤ 35.

  • Number of product attributes is ≥ 2 and ≤ 8.

  • Number of consumer attributes is ≤ 5.

  • Number of advertising attributes is ≤ 2.

The attributes need not be standardized (the code will standardize them internally). However, we do recommend de-meaning the attributes which makes the intercepts (\(\alpha_0, \eta_0\)) easier to interpret.

In addition, the NNE is pretrained on data with following characteristics. If your data’s characteristics fall outside these ranges, the pretrained NNE may not work as intended.

  • Buy rate (fraction of consumers who bought inside good) is 0.5% ~ 70%.

  • Search rate (fraction of consumers who went beyond free search) is 1% ~ 80%.

  • Average number of searches per consumer is 1 ~ 6.

Description of files#

trained_nne.mat#

This file stores the trained neural net as well as pre-defined settings.

  • nne: a structure that stores the trained neural net, as well as some settings used in training, such as the range of data dimensions and the prior of search model parameters.

nne_estimate.m#

This is the main function that applies pretrained NNE to your data.

result = nne_estimate(nne, Y, Xp, Xa, Xc, consumer_idx, se = false, checks = true)

Inputs:

  • nne: as described above, from trained_nne.mat.

  • Y: \(nJ\) by 2 matrix. The \(((i-1)J+j)\)th row corresponds to product \(j\) for consumer \(i\). The first value indicates if the product is searched. The second value indicates if the product is bought.

  • Xp: a matrix with \(nJ\) rows. The \(((i-1)J+j)\)th row stores the product attributes of product \(j\) for consumer \(i\).

  • Xa: a matrix with \(nJ\) rows. The \(((i-1)J+j)\)th row stores the advertising attributes of product \(j\) for consumer \(i\).

  • Xc: a matrix with \(n\) rows. The \(i\)th row stores the consumer attributes of consumer \(i\).

  • consumer_idx: a column vector with \(nJ\) values. The \(((i-1)J+j)\)th value equals \(i\).

  • se = false: optional input. Set it to “true” to calculate bootstrap standard errors.

  • checks = true: optional input. Set it to “false” to omit all sanity checks.

Outputs:

  • result: a table showing parameter estimate and bootstrap standard errors (if se = true).

moments.m#

This function is used by nne_estimate.m. It computes moments from data. The moments include summary statistics as well as coefficients of reduced-form regressions. The moments are fed into the neural net as inputs.

reg_logit.m#

This function is used by moments.m. It performs ridge logit or multinomial-logit regressions. It runs faster than Matlab built-in regressions and this has sped up pretraining. This faster execution is less important when applying the pretrained NNE. Nevertheless, we continue to make use of it.

reg_linear.m#

This function is used by moments.m. It performs ridge linear regression.

data_checks.m#

This function is used by nne_estimate.m to run some basic sanity checks on data. For example, did every consumer buy at most one option; did every consumer conduct the free search?


Note: The following files are not required to apply the pretrained NNE. They are provided for reference purposes.

winsorize.m#

This function winsorizes data at 0.5 and 99.5 percentiles. We suggest applying it to your data before applying the pretrained NNE. For example, Xp = winsorize(Xp).

curve.mat#

This file stores a lookup table for the relation between search cost and reservation utility. This relation is useful when we compute the optimal consumer choices in the sequential search model.

search_model.m#

This function codes the search model. The function is used in pretraining but not in application. We provide it for reference purpose. It is also useful if you want to conduct Monte Carlo experiments.

Y = search_model(par, curve, Xp, Xa, Xc, consumer_idx)

Inputs:

  • par: vector of the parameter value for the search model.

  • curve: as described above, from curve.mat.

  • Xp, Xa, Xc, and consumer_idx: data formatted as described before.

Outputs

  • Y: a matrix with the simulated search and purchase choices, formatted as described before.