Configuração de estratégia de momentum e carry para pares de FX

# setup.yaml


```yaml
strategy_id: fx_pairs
setup_version: v1

universe:
  symbols:
    - AUD_JPY
    - AUD_NZD
    - AUD_USD
    - CAD_JPY
    - CHF_JPY
    - EUR_AUD
    - EUR_CAD
    - EUR_CHF
    - EUR_GBP
    - EUR_JPY
    - EUR_USD
    - GBP_AUD
    - GBP_CHF
    - GBP_JPY
    - GBP_USD
    - NZD_JPY
    - NZD_USD
    - USD_CAD
    - USD_CHF
    - USD_JPY
  n_assets: 20

decision:
  cadence: daily_ny_close
  snapshot: ny_5pm_close
  execution_delay: next_bar_open
  # The venue calendar that implements the 5PM rollover, and so the one that assigns a
  # four-hour bar to the session it was printed in. Distinct from `evaluation.calendar`,
  # which is the calendar the cross-validation splitter counts train and validation
  # windows on. `02_labels` and `04_model_based_features` aggregate on this same value.
  session_calendar: CME_FX

# Engine-level execution defaults. Single source of truth — Ch16-19 notebooks
# read these via get_backtest_config(); never declare a local INITIAL_CASH
# or share_type constant. Changing values here invalidates every existing
# backtest_hash for this case study (cash + share_type are spec-hash inputs).
#
# ``allocator_lookback`` is the bars-of-underlying-price window applied
# uniformly to every moment-based allocator (inverse_vol, risk_parity, hrp,
# mvo_ledoit_wolf). FX pairs has daily OANDA bars → 63 ≈ 3 months. Keeps
# allocators comparable — MVO is not granted a longer covariance window than
# IV by default.
execution:
  initial_cash: 100_000          # IBKR retail cohort default
  share_type: integer            # FX traded in units of base currency, integer share semantics
  allocator_lookback: 63         # 3 months of daily OANDA bars

mapping:
  class: long_short_rank_rebalance
  position_state_space: long_short
  entry_logic: rank_by_momentum_or_carry
  sizing: equal_weight

costs:
  class: material
  components: [spread, swap_points]
  spread_bps:
    major_pairs: [1, 3]
    cross_pairs: [3, 8]

backtest:
  rebalance:
    # A rebalance is skipped when the per-asset weight change is below
    # min_weight_change AND the resulting trade notional is below
    # min_trade_value. The benchmark profile disables thresholds so that
    # full-universe equal-weight (1/N per asset) rebalances at all.
    default:
      min_weight_change: 0.005
      min_trade_value: 100.0
    benchmark:
      min_weight_change: 0.0
      min_trade_value: 0.0
  sweep:
    # Iteration controls per stage. ``signal: 0`` means "all predictions";
    # downstream stages take the top-N from the upstream stage's rank-1.
    # Notebooks read these via get_top_n_predictions(case_study, stage).
    top_n_predictions:
      signal: 0                 # all signal predictions (eq-weight baseline)
      allocation: 10            # top-10 model configs by equal-weight baseline Sharpe
      cost_sensitivity: 1       # top-1 of {signal+allocation} per label
      risk_overlay: 1           # top-1 of {signal+allocation} per label
    # Skip MVO/HRP when allocator runtime is the bottleneck (intraday CSes).
    expensive_allocators_skip: false
    # All alternative allocators take the cheap path except the two moment-based
    # methods covered by expensive_allocators_skip. Equal weight is the baseline
    # and is intentionally absent here so the allocation stage cannot double-count it.
    # Ch16 baseline selection. Long-short equal-weight top-k on all
    # three label horizons; long-short is enforced at the backtest config
    # level (allow_short_selling=true).
    top_k_grid:
      fwd_ret_1d:  [5, 10]
      fwd_ret_5d:  [5, 10]
      fwd_ret_21d: [5, 10]
    # Ch17 portfolio: reuses top_k_grid above and sweeps over allocators.
    # Moment-based allocators (IV/RP/HRP/MVO_LW) all use the CS-level
    # ``execution.allocator_lookback`` — no per-allocator vol_window / lookback
    # fields. No max-weight cap: each allocator expresses its natural
    # concentration profile so the cross-method comparison is honest.
    allocators:
      - {name: score_weighted,  method: score_weighted}
      - {name: inverse_vol,     method: inverse_vol}
      - {name: risk_parity,     method: risk_parity}
      - {name: mvo_ledoit_wolf, method: mvo_ledoit_wolf}
      - {name: hrp,             method: hrp}
      - {name: conformal_weighted, method: conformal_weighted}
    # Ch18 cost sensitivity (bps; spread is the dominant cost component,
    # captured here as combined commission + slippage in bps).
    cost_grid_bps: [0, 1, 2, 3, 5, 7, 10, 15, 20, 30, 50]
    # Ch19 risk overlays.
    risk_controls:
      position:
        - {name: stop_loss_3pct,  type: stop_loss,     threshold: 0.03}
        - {name: stop_loss_5pct,  type: stop_loss,     threshold: 0.05}
        - {name: stop_loss_10pct, type: stop_loss,     threshold: 0.10}
        - {name: stop_loss_15pct, type: stop_loss,     threshold: 0.15}
        - {name: trailing_1pct,   type: trailing_stop, threshold: 0.01}
        - {name: trailing_2pct,   type: trailing_stop, threshold: 0.02}
        - {name: trailing_3pct,   type: trailing_stop, threshold: 0.03}
        - {name: trailing_5pct,   type: trailing_stop, threshold: 0.05}
        - {name: trailing_10pct,  type: trailing_stop, threshold: 0.10}
        - {name: trailing_15pct,  type: trailing_stop, threshold: 0.15}
        - {name: trailing_20pct,  type: trailing_stop, threshold: 0.20}
        - {name: time_exit_10,    type: time_exit,     bars: 10}
        - {name: time_exit_20,    type: time_exit,     bars: 20}
        - {name: time_exit_40,    type: time_exit,     bars: 40}

# The feature-specification register. ``03_financial_features`` binds every window
# from here and renders ``families`` as the register table; nothing in that notebook
# retypes a number this block declares. ``lookback`` is the longest chain of trailing
# bars a family reads, counted from the decision timestamp, and it is the floor the
# warmup audit holds each column to. ``lag`` is zero throughout: every input is an
# OANDA spot bar, on the tape at the NY 5PM snapshot the decision is taken at.
features:
  windows:
    momentum: [5, 10, 21, 42, 63, 126, 252]
    close_to_close_volatility: [21, 63]
    garman_klass: [21, 63, 126, 252]
    zscore: 252              # trailing window the multi-horizon returns are standardized against
    zscore_horizons: [21, 63, 126]
    channel: [21, 63, 126]
    skip_recent: 21
    drawdown: [63]
    rsi: [14, 63]
    moving_average: [21, 63, 252]
    range: [21, 63]
    bollinger: 21
    dollar: [21, 63]         # horizons the dollar proxy is accumulated over
    dollar_exposure: 252     # window each pair's correlation with the proxy is taken over
  ranked:
    - ret_21d
    - ret_63d
    - ret_126d
    - vol_gk_21d
    - vol_gk_63d
    - sharpe_21d
    - sharpe_63d
    - sharpe_126d
  # Rows are kept from the bar this column can first hold a value, which is the
  # longest chain in the matrix and the point past which every family is dense.
  null_policy_carrier: zscore_126d
  persistence_horizon: 63    # bars F6 reads the feature autocorrelation out to
  # Rank correlation at which the redundancy tree is cut: two columns agreeing this
  # closely on their ordering are one column as far as a linear model is concerned.
  redundancy_cut: 0.7
  families:
    - name: momentum
      pattern: ret_*d|mom_skip_recent|accel_*
      role: signal
      hypothesis: A pair that has moved keeps moving over the following session
      inputs: daily close
      lookback: 252
      lag: 0
      frame: time series
      representation: simple return, and differences between horizons
      failure_mode: reverses at the shortest horizons
    - name: risk-adjusted momentum
      pattern: sharpe_*d
      role: signal
      hypothesis: Trend earned with less dispersion repeats more reliably
      inputs: daily log return
      lookback: 252
      lag: 0
      frame: time series
      representation: mean log return over its own dispersion, annualized
      failure_mode: unbounded when dispersion approaches zero
    - name: mean reversion
      pattern: zscore_*|channel_pos_*|bollinger_pctb_*
      role: signal
      hypothesis: A pair stretched against its own recent range comes back
      inputs: daily close
      lookback: 378
      lag: 0
      frame: time series
      representation: trailing z-score, position in the trailing range
      failure_mode: a trending pair sits at the edge of its channel for months
    - name: volatility
      pattern: vol_gk_*|vol_cc_*|vol_ratio_*
      role: state
      hypothesis: Dispersion sets how far apart the cross-section can spread
      inputs: daily OHLC
      lookback: 252
      lag: 0
      frame: time series
      representation: Garman-Klass and close-to-close deviation, and ratios of windows
      failure_mode: lags a shock by roughly half its window
    - name: range and drawdown
      pattern: max_dd_*|avg_range_*
      role: state
      hypothesis: How far a pair sits below its peak conditions what a signal is worth
      inputs: daily OHLC
      lookback: 63
      lag: 0
      frame: time series
      representation: share below the trailing peak, mean normalized daily range
      failure_mode: says nothing about direction
    - name: oscillator and trend
      pattern: rsi_*|price_to_ma_*
      role: signal
      hypothesis: Price against its own recent path separates trend from exhaustion
      inputs: daily close
      lookback: 252
      lag: 0
      frame: time series
      representation: Wilder-smoothed oscillator, ratio to a moving average
      failure_mode: saturates in a sustained trend
    - name: dollar factor
      pattern: usd_factor_*|usd_corr_*
      role: state
      hypothesis: A broad dollar move is common to seven pairs and is not pair-specific
      inputs: daily returns of the seven USD pairs
      lookback: 315
      lag: 0
      frame: cross-asset
      representation: signed average return, and each pair's rolling correlation with it
      failure_mode: a signed average is not an estimated factor
    - name: cross-sectional position
      pattern: rank_*
      role: signal
      hypothesis: Only relative standing is tradable in a long-short rank strategy
      inputs: ret_21d, ret_63d, ret_126d, vol_gk_21d, vol_gk_63d, sharpe_21d, sharpe_63d, sharpe_126d
      lookback: 252
      lag: 0
      frame: cross-section
      representation: percentile within the decision date
      failure_mode: discards the level the state families carry instead

# What `04_model_based_features` decides, declared here for the same reason the
# feature windows above are: an estimation window is part of a fitted feature's
# definition, so it belongs where the definition lives rather than inside the
# notebook that runs it. Every count is in trading sessions on the FX calendar.
#
# A fitted feature is bounded by the schedule below and not by a cross-validation
# fold. The parameters behind a value are estimated from sessions strictly before
# it, refreshed on the cadence given, and the same value comes out whichever fold
# later selects the row - so the artifact carries no fold column.
#
# The panel runs 2011-01-03 to 2025-12-31, 3,874 sessions, and every pair quotes
# on all of them. Only one session precedes the oldest fold's training start, so
# each burn-in below is paid out of that fold's training window rather than ahead
# of it. The sessions it costs are named against each entry; no validation window
# is touched, because the earliest one opens 2016-01-05.
model_based:
  kalman:
    # Sessions of a pair's own log price before its local linear trend model is
    # fitted. One year. The three noise variances are separated by how often the
    # level moves against how far a quote strays from it, and a shorter window
    # does not contain enough of either to tell them apart. It costs the oldest
    # fold 251 of its 1,289 training sessions and no validation session.
    burnin: 252
    # How often the three variances are re-estimated. A quarter. Each estimate is
    # a Nelder-Mead search over the whole expanding history, which is the most
    # expensive fit in the notebook, and quoting noise is not a fast-moving
    # quantity. The first value lands 2011-12-22.
    refit_every: 63
    # Iteration cap for that search.
    maxiter: 300
  hmm:
    # A calm dollar state and a turbulent one.
    n_states: 2
    # Sessions of the dollar factor before the regime chain is fitted. Two years.
    # A two-state chain has to see both states switch several times before its
    # transition matrix means anything, and this is one market-level series
    # rather than a panel, so the burn-in is paid once for the whole notebook.
    # It matches cme_futures, whose regime model reads a daily market-level
    # series of the same shape. The dollar factor itself starts 2011-02-01,
    # after the 21-session volatility window it carries, so the first regime
    # value lands 2013-01-15: 504 of the oldest fold's 1,269 dollar-factor
    # training sessions, and no validation session.
    burnin: 504
    # How often the chain is re-estimated. A quarter, as in etfs and cme_futures:
    # regime parameters are the slowest-moving thing this notebook fits and each
    # estimate costs `n_restarts` expectation-maximization searches.
    refit_every: 63
    # Expectation-maximization reaches a local optimum, so each fit is repeated
    # from this many starting points and the highest training likelihood is kept.
    n_restarts: 10
    # A restart is rejected when its final EM step falls by more than this
    # fraction of the log-likelihood's own magnitude. Real divergence moves
    # hundreds of nats; what this has to tolerate is single digits against a
    # likelihood of about 4.3e4.
    stability_rel_tol: 0.001
  arima:
    # Sessions of a pair's own returns before its short-memory return model is
    # fitted. One year, matching cme_futures' ARIMA, which is the same kind of
    # model on the same kind of daily per-entity series. A return needs the
    # session before it, so this series starts one session later than the price
    # panel and the first value lands 2011-12-23, costing the oldest fold 252 of
    # its 1,289 training sessions and no validation session.
    burnin: 252
    # Monthly. The coefficients of a one-lag return model are the fastest-moving
    # parameters here and the fit is cheap, so this is the shortest cadence in
    # the block.
    refit_every: 21
    # One autoregressive term and one moving-average term - the shortest memory
    # the family offers, which is all daily currency returns support.
    order: [1, 0, 1]

evaluation:
  n_splits: 8
  train_size: P5Y
  val_size: P1Y
  holdout_start: '2024-01-01'
  holdout_end: '2025-12-31'
  calendar: FX
  periods_per_year: 252  # FX 5d/wk

labels:
  primary: fwd_ret_1d
  buffer: 1D
  variants:
    - fwd_ret_5d
    - fwd_ret_21d
  variant_buffers:
    fwd_ret_5d: 5D
    fwd_ret_21d: 21D
  # Vectorized-backtest thinning step per label: number of schedule slots
  # to advance per trade so holding periods don't overlap.
  rebalance_step:
    fwd_ret_1d: 1
    fwd_ret_5d: 5
    fwd_ret_21d: 21

modeling:
  gbm:
    libraries: [lightgbm]
    preset: default
    device: cpu
    # LightGBM's own CPU default. 63 is the GPU default and was carried over with the
    # device when these runs moved off the GPU, so every CPU fit was quantizing the design
    # matrix into a quarter of the bins the library would have used. Coarser bins are
    # faster and lose split points; the reader running this on a CPU gets what the
    # documentation describes.
    max_bin: 255

causal:
  treatment: mom_skip_recent
  # Bars the treatment's own construction window spans, which is what the placebo block has
  # to cover: permuting mom_skip_recent in blocks shorter than this destroys the serial
  # dependence the refutation exists to preserve, and the resulting p-value reads like a
  # refutation without being one. Declared here rather than inferred, because guessing which
  # element of a window list a column was built from puts a wrong number behind a right-looking
  # one. Derived from the construction, not chosen:
  #
  # `close.shift(skip_recent) / close.shift(momentum[-1]) - 1` in 03_financial_features,
  # which is 21 and 252. Three different numbers live in that one expression and conflating them
  # is easy, so all three are stated: the LOOKBACK is 252 sessions, the oldest price it reads;
  # the RETURN INTERVAL is 231 sessions, from t-252 to t-21, which is what the value measures;
  # and consecutive values OVERLAP in 230 of those 231. The declaration takes the lookback
  # because it is the larger, so a block of this length spans the 231-session dependence
  # whichever way the arithmetic is read.
  treatment_window: 252
  confounders: [vol_gk_21d, vol_gk_63d, zscore_21d]
  method: walk_forward_dml
  # Two things the register above cannot supply, and this is why the number is written here
  # rather than derived. `features.windows.momentum` is a list of bar counts, not a
  # suffix-keyed mapping, and `mom_skip_recent` is not built from any single element of it.
  #
  # And the block has to span this window, which is the reason it is declared at all. A block
  # permutation keeps the treatment's serial dependence while breaking its alignment with the
  # outcome; a block shorter than the construction window breaks the dependence too, which
  # narrows the placebo distribution and makes the p-value read stronger than the evidence is.
  # Sized from the label buffer alone this case study permuted a 252-session momentum column
  # in blocks of 1, 5 and 21.

```