ML Strategies with PortfolioTesteR (Advanced Getting Started)"
In PortfolioTesteR: Test Investment Strategies with English-Like Code

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 7,
  fig.height = 5,
  dpi = 96,
  message = FALSE,
  warning = FALSE
)
set.seed(1)

# CRAN vs local
heavy <- identical(Sys.getenv("NOT_CRAN"), "true")  # CRAN sets this to "false"

# Optional backends (Suggests)
has_xgb <- requireNamespace("xgboost", quietly = TRUE)
has_rf  <- requireNamespace("ranger",  quietly = TRUE)

# Sequences are OFF by default (CRAN-safe). Opt-in locally with:
# Sys.setenv(RUN_SEQ = "true")  # or RUN_KERAS_VIGNETTES = "true" (back-compat)
run_seq <- FALSE
run_seq_env <- tolower(Sys.getenv("RUN_SEQ", Sys.getenv("RUN_KERAS_VIGNETTES", "false")))
if (heavy && run_seq_env %in% c("true","1","yes") && requireNamespace("keras", quietly = TRUE)) {
  # only now ask Keras if it's actually usable (this can spin up TF)
  run_seq <- isTRUE(try(keras::is_keras_available(), silent = TRUE))
}

# Base-R helper to sandbox TF Autograph cache per-chunk (no 'withr' usage)
tf_autograph_scope <- function() {
  ag_dir <- file.path(tempdir(), paste0("tf_ag_cache_", as.integer(runif(1, 1, 1e9))))
  dir.create(ag_dir, showWarnings = FALSE, recursive = TRUE)
  old_ag <- Sys.getenv("TF_AUTOGRAPH_CACHE_DIR", unset = NA_character_)
  Sys.setenv(TF_AUTOGRAPH_CACHE_DIR = ag_dir)
  # Ensure cleanup when the chunk completes
  on.exit({
    if (!is.na(old_ag)) Sys.setenv("TF_AUTOGRAPH_CACHE_DIR", old_ag) else Sys.unsetenv("TF_AUTOGRAPH_CACHE_DIR")
    unlink(ag_dir, recursive = TRUE, force = TRUE)
    unlink(Sys.glob(file.path(tempdir(), "__autograph_generated_file*.py")), force = TRUE)
    unlink(file.path(tempdir(), "__pycache__"), recursive = TRUE, force = TRUE)
  }, add = TRUE)
  invisible(NULL)
}

Why ML in PortfolioTesteR

PortfolioTesteR lets you test ML-based stock selection with the same workflow you already know:

1) Prepare lagged, leakage-safe features + future labels
2) Fit -> Predict on a rolling IS/OOS schedule
3) Transform -> Select -> Weight the scores
4) Backtest and inspect diagnostics

All examples use the bundled weekly/daily datasets; no internet is required.

suppressPackageStartupMessages(library(PortfolioTesteR))
suppressPackageStartupMessages(library(data.table))

# Bundled sample data
data(sample_prices_weekly)
data(sample_prices_daily)

# Optional peek
head(sample_prices_weekly[, 1:6])

1) Baseline ML (pooled): Ridge regression

Simple pooled regression on tabular features. The helpers build lagged features and future labels (4-week horizon).

# Prepare tabular features and labels
X <- ml_prepare_features(sample_prices_weekly, sample_prices_daily)
Y <- make_labels(sample_prices_weekly, horizon = 4L, type = "log")

# Linear/Ridge baseline
ridge <- ml_make_model("ridge")

set.seed(1)
res_baseline <- ml_backtest(
  features_list = X, labels = Y,
  fit_fn  = ridge$fit,
  predict_fn = ridge$predict,
  schedule = list(is = 104L, oos = 4L, step = 4L),  # 2y IS, 1m OOS, monthly step
  transform = "zscore",                              # IS-only scaling (no leakage)
  selection = list(top_k = 15L),
  weighting = list(method = "rank"),
  prices = sample_prices_weekly,
  initial_capital = 1e5,
  name = "Baseline Ridge (pooled)"
)

print(res_baseline$backtest)
plot(res_baseline$backtest, type = "performance")

2) Ensemble: Ridge + XGBoost

This section is optional. It runs only if at least one of ranger/xgboost is installed.

# Start with ridge (always available)
models <- list(ml_make_model("ridge"))

# Add RF only if ranger is installed
if (has_rf) {
  models <- c(models, list(ml_make_model("rf", num.trees = if (heavy) 400L else 200L, mtry = 3L)))
}

# Add XGBoost only if xgboost is installed
if (has_xgb) {
  models <- c(models, list(
    ml_make_model(
      "xgboost",
      params  = list(objective = "reg:squarederror", max_depth = 4L, eta = 0.07),
      nrounds = if (heavy) 300L else 150L
    )
  ))
}

# Build the ensemble from whatever is available
ens <- do.call(ml_make_ensemble, models)

set.seed(2)
res_ens <- ml_backtest(
  features_list = X, labels = Y,
  fit_fn = ens$fit, predict_fn = ens$predict,
  schedule = list(is = 104L, oos = 4L, step = 4L),
  transform = "zscore",
  selection = list(top_k = 15L),
  weighting = list(method = "rank"),
  prices = sample_prices_weekly,
  initial_capital = 1e5,
  name = "Ensemble (available learners)"
)

print(res_ens$backtest)
plot(res_ens$backtest, type = "performance")

3) Sector-neutral modelling (tabular XGBoost, per-group)

This section is optional and runs only if xgboost is installed.

symbols <- setdiff(names(sample_prices_weekly), "Date")
gmap    <- demo_sector_map(symbols, n_groups = 4L)  # demo mapping for the sample data

xgb_g <- ml_make_model(
  "xgboost",
  params  = list(objective = "reg:squarederror", max_depth = 3L, eta = 0.05),
  nrounds = if (heavy) 250L else 150L
)

set.seed(3)
res_xgb_sect <- ml_backtest(
  features_list = X, labels = Y,
  fit_fn = xgb_g$fit, predict_fn = xgb_g$predict,
  schedule = list(is = 104L, oos = 4L, step = 4L),
  group = "per_group", group_map = gmap,   # sector-neutral training
  transform = "zscore",
  selection = list(top_k = 15L),
  weighting = list(method = "softmax", temperature = 12),
  prices = sample_prices_weekly,
  initial_capital = 1e5,
  name = "XGBoost (per-sector neutral)"
)

print(res_xgb_sect$backtest)
plot(res_xgb_sect$backtest, type = "performance")

4) Sequence model (GRU, pooled): "returns pyramid"

Optional and disabled by default. Enable locally by running Sys.setenv(RUN_SEQ = "true") before knitting, and only if your installation supports sequence models.

tf_autograph_scope()  # isolate TF Autograph cache to avoid detritus NOTE

# Build a compact 'returns pyramid' of momentum-like sequences (lagged)
r1  <- panel_lag(calc_momentum(sample_prices_weekly,  1L), 1L)
r4  <- panel_lag(calc_momentum(sample_prices_weekly,  4L), 1L)
r12 <- panel_lag(calc_momentum(sample_prices_weekly, 12L), 1L)

features_seq <- list(r1 = r1, r4 = r4, r12 = r12)
Y_seq <- make_labels(sample_prices_weekly, horizon = 4L, type = "log")

units  <- if (heavy) 32L else 8L
epochs <- if (heavy) 12L else 4L

seq_gru <- ml_make_seq_model(
  "gru",
  steps  = 26L,
  units  = units,
  epochs = epochs,
  seed   = 42L
)

set.seed(42)
res_seq_gru <- ml_backtest_seq(
  features_list   = features_seq,
  labels          = Y_seq,
  steps           = 26L,
  horizon         = 4L,
  fit_fn          = seq_gru$fit,
  predict_fn      = seq_gru$predict,
  schedule        = list(is = 104L, oos = 4L, step = 4L),
  group           = "pooled",
  normalize       = "zscore",
  selection       = list(top_k = 10L),
  weighting       = list(method = "softmax", temperature = 12),
  prices          = sample_prices_weekly,
  initial_capital = 1e5,
  name            = "Seq-GRU (pooled): returns pyramid (26x3)"
)

print(res_seq_gru$backtest)
plot(res_seq_gru$backtest, type = "performance")

5) Sequence model (GRU, per-group): sector-neutral sequences

Optional and disabled by default. Enable with RUN_SEQ="true".

tf_autograph_scope()  # isolate TF Autograph cache for this chunk too

symbols <- setdiff(names(sample_prices_weekly), "Date")
gmap    <- demo_sector_map(symbols, n_groups = 4L)

units_g  <- if (heavy) 32L else 8L
epochs_g <- if (heavy) 12L else 4L

seq_gru_g <- ml_make_seq_model(
  "gru",
  steps  = 26L,
  units  = units_g,
  epochs = epochs_g,
  seed   = 123L
)

set.seed(123)
res_seq_gru_g <- ml_backtest_seq(
  features_list   = features_seq,
  labels          = Y_seq,
  steps           = 26L,
  horizon         = 4L,
  fit_fn          = seq_gru_g$fit,
  predict_fn      = seq_gru_g$predict,
  schedule        = list(is = 104L, oos = 4L, step = 4L),
  group           = "per_group", group_map = gmap,
  normalize       = "zscore",
  selection       = list(top_k = 10L),
  weighting       = list(method = "softmax", temperature = 12),
  prices          = sample_prices_weekly,
  initial_capital = 1e5,
  name            = "Seq-GRU (per-sector neutral): returns pyramid (26x3)"
)

print(res_seq_gru_g$backtest)
plot(res_seq_gru_g$backtest, type = "performance")

6) Light tuning: Top-K and softmax temperature

Runs only off-CRAN (when NOT_CRAN=true).

topk_vals <- c(8L, 10L, 12L, 15L)
temp_vals <- c(8, 12, 16)

score_tbl <- tune_ml_backtest(
    features_list = X, labels = Y, prices = sample_prices_weekly,
    fit_fn = ridge$fit, predict_fn = ridge$predict,
    schedule = list(is = 104L, oos = 4L, step = 4L),
    grid = list(
        top_k = topk_vals,
        temperature = temp_vals,
        method = "softmax",
        transform = "zscore"
    )
)

score_tbl[order(-sharpe)][1:10]

Recap & Next Steps

You saw pooled vs per-group (sector-neutral) training,
Tabular vs sequence learners, and
A small tuning exercise reusing the same IS/OOS schedule.

Try next:
1) Replace GRU with LSTM/CNN1D if available (still gated by RUN_SEQ).
2) Blend tabular and sequence scores (e.g., average ranks) before weighting.
3) Add exposure caps or per-sector position limits when selecting Top-K.

# Best-effort cleanup of TensorFlow Autograph cache files that can trigger
# "detritus in the temp directory" NOTE
try(unlink(Sys.glob(file.path(tempdir(), "__autograph_generated_file*.py")), force = TRUE), silent = TRUE)
try(unlink(file.path(tempdir(), "__pycache__"), recursive = TRUE, force = TRUE), silent = TRUE)