LoginSign Up
← Back to all sparks
L

lares

INFRA · APIS
Velocity0.0

Lean Analytics and Robust Exploration Sidekick

github.com/laresbernardo/lares ↗

One analyst's toolbox, still shipping: Meta MMM tooling beside Wordle solvers

marketing-mix-modelingrobynutility-toolboxapi-compatibilityautoml
◆Current state
lares is Bernardo Lares' personal R sidekick and has never pretended otherwise — the same release adds a marketing-mix diagnostic and a function to keep your screen awake. The serious core is the Robyn/Meta arm: model selection, per-channel performance decomposition, cross-channel budget allocation. Around it sit h2o AutoML wrappers, credential and cache helpers, MP3 tagging, and a family of games.
◆Where it's heading
Over two years the Robyn tooling matured first — robyn_performance() and robyn_modelselector() were rebuilt release after release through 2024 — and recent work has drifted outward into media and API plumbing: encrypted file helpers in 5.3.2, then an iTunes-first metadata path, MP3 tag writing, and a move of holidays() onto the Nager.Date API in 5.4.0. Cadence is roughly quarterly and the package has been quiet for three months.
◆Prediction
The pattern in nearly every release is that a third-party API moves and lares follows it — Meta's version bumps, Robyn 3.12.0, now Nager.Date. Expect the next release to be another compatibility pass plus whatever utility the author needed that week; nothing in these entries points at a structural change.

◆Recent moves

  1. 3mo ago

    Nager.Date holidays, MP3 tag writing, named h2o models

    The broadest release in the window and still recognisably a grab bag: holidays() moves to the Nager.Date API with ISO country codes, mp3_get() prefers iTunes over Spotify for metadata, mp3_update_tags() writes tags back, and h2o_automl() gains model_name so exported artefacts stop colliding. Splitting other_functions.R into three files is the first real housekeeping the package has done.

    View source ↗
  2. 11mo ago

    Encrypted file helpers join the credential tooling

    encrypt_file(), read_encrypted() and write_encrypted() extend the existing credential and cache machinery to data at rest, which matters for a package whose day job includes holding API keys for Meta and OpenAI. The remainder is connection handling and message ordering.

    View source ↗
  3. 1y ago

    ggplot2 4.0 readiness and Meta API v21

    Compatibility work dominates: preparing for ggplot2 4.0.0, moving the Meta API from v20 to v21, and switching get_mp3() to yt-dlp. Reweighted confidence criteria in robyn_modelselector() are the one substantive change to the marketing arm.

    View source ↗
  4. 1y ago

    Marginal CPA inverted and multi-model performance fixed

    A correctness release for the Robyn arm — the mCPA calculation was inverted, and robyn_performance() mishandled OutputCollect objects carrying more than one model. Either would have quietly produced wrong channel economics.

    View source ↗
  5. 1y ago

    Certainty and cluster-spread metrics for model selection

    robyn_modelselector() gains certainty and cluster_sd scores, continuing the slow build-out of automated candidate ranking that is this package's most developed idea. normalize() gains a custom scale range and robyn_performance() can include non-promotional contributions.

    View source ↗
  6. 1y ago

    Cross-channel budget allocation across several models

    robyn_xchannels() extends budget allocation past a single model, which is the direction the Robyn tooling has been heading — from scoring one model to reasoning over a set of them. The rest is number formatting and small predicates.

    View source ↗