ml_backtest_seq: One-call backtest wrapper (sequence features)

In PortfolioTesteR: Test Investment Strategies with English-Like Code

One-call backtest wrapper (sequence features)

Description

One-call backtest wrapper (sequence features)

Usage

ml_backtest_seq(
  features_list,
  labels,
  steps = 26L,
  horizon = 4L,
  fit_fn,
  predict_fn,
  schedule = list(is = 156L, oos = 4L, step = 4L),
  group = c("pooled", "per_symbol", "per_group"),
  group_map = NULL,
  normalize = c("none", "zscore", "minmax"),
  selection = list(top_k = 15L, max_per_group = NULL),
  weighting = list(method = "softmax", temperature = 12, floor = 0),
  caps = list(max_per_symbol = 0.1, max_per_group = NULL),
  prices,
  initial_capital = 1e+05,
  name = "SEQ backtest"
)

Arguments

features_list	list of panels to be stacked over steps history.
labels	future-return panel aligned to the features.
steps	int; lookback length (e.g., 26).
horizon	int; label horizon (e.g., 4).
fit_fn	function ⁠(X, y) -> model⁠ trained on in-sample stacked rows.
predict_fn	function ⁠(model, Xnew) -> numeric⁠ scores.
schedule	list with elements is, oos, step.
group	one of "pooled", "per_symbol", "per_group".
group_map	optional data.frame(Symbol, Group) if group = "per_group".
normalize	"none", "zscore", or "minmax" applied using IS data only.
selection	list: top_k, max_per_group (optional).
weighting	list: method, temperature, floor.
caps	list: max_per_symbol, optionally max_per_group.
prices	price panel used by the backtester (Date + symbols).
initial_capital	starting capital.
name	string for the backtest result.

Value

list: scores, mask, weights, backtest.

Examples

# as above, but with steps/horizon and normalize

PortfolioTesteR documentation built on Nov. 5, 2025, 5:23 p.m.