Consider N best models to select the right ones to study using several criteria/metrics such as potential improvement on budget allocator, how many non-zero coefficients there are, R squared, historical performance, baseline expectation, etc.
Read more about this functionality in Medium post: here.
robyn_modelselector(
InputCollect,
OutputCollect,
metrics = c("rsq_train", "performance", "potential_improvement", "non_zeroes",
"incluster_models", "cluster_sd", "certainty", "baseline_dist"),
wt = c(2, 0.5, 0, 1, 0.1, 0, 0, 0),
baseline_ref = 0,
top = 4,
n_per_cluster = 5,
allocator_limits = c(0.5, 2),
quiet = FALSE,
cache = TRUE,
...
)# S3 method for robyn_modelselector
plot(x, ...)
list with resulting ranked data.frames, weights and plot.
Robyn output objects.
Character vector. Which metrics do you want to consider?
Pick any combination from: "rsq_train" for trained R squared,
"performance" for ROAS or (inverse) CPA, "potential_improvement" for
default budget allocator improvement using allocator_limits,
"non_zeroes" for non-zero beta coefficients, "incluster_models" for
amount of models per cluster, "baseline_dist" for the difference between
the model's baseline and baseline_ref value, "certainty" metric
to minimize the channels' distance to their cluster's mean performance,
weighted by spends spend_wt = TRUE, "cluster_sd" metric to score
based on the paid channels' performance standard deviations in clusters.
Additionally, you can use the standard MOO errors:
"nrmse", "decomp.rssd", and "mape" (the lowest the error, the highest
the score; same for "baseline_dist" and "cluster_sd").
Vector. Weight for each of the normalized metrics selected,
to calculate the score and rank models. Must have the same order and length
of metrics parameter input.
Numeric value. Between 0 and 1. What is the baseline percentage you expect? Baseline in this case are all the sales or conversions from non-media channels (organic & paid). Use with "baseline_dist" metric.
Integer. How many ranked models to star? The better the model is, the more stars it will have marked.
Integer. How many models per cluster do you want to plot? Default: 5. Keep in mind they will all be considered for the calculations.
Numeric vector, length 2. How flexible do you want to be with the budget allocator? By default, we'll consider a 0.5X and 2X range to let the budget shift across channels.
Boolean. Keep quiet? If not, message will be shown.
Use cache functionality for allocator's results?
Additional parameters.
robyn_modelselector object
Other Robyn:
robyn_hypsbuilder(),
robyn_marginal(),
robyn_performance()