Concevoir un cadre de recherche diversifié multi-actifs sur les contrats à terme CME

# setup.yaml


```yaml
strategy_id: cme_futures
setup_version: v1

# 30 CME products grouped by sector. Tickers are CME root symbols.
universe:
  product_groups:
    equity_index:
      - ES   # E-mini S&P 500
      - NQ   # E-mini Nasdaq-100
      - YM   # E-mini Dow Jones Industrial Average
      - RTY  # E-mini Russell 2000
    treasuries:
      - ZN   # 10-year T-Note
      - ZB   # 30-year T-Bond
      - ZF   # 5-year T-Note
      - ZT   # 2-year T-Note
    energy:
      - CL   # WTI Crude Oil
      - NG   # Henry Hub Natural Gas
      - HO   # NY Harbor ULSD (heating oil)
      - RB   # RBOB Gasoline
    metals:
      - GC   # Gold
      - SI   # Silver
      - HG   # Copper
      - PL   # Platinum
    currencies:
      - 6E   # Euro / US Dollar
      - 6J   # Japanese Yen / US Dollar
      - 6B   # British Pound / US Dollar
      - 6A   # Australian Dollar / US Dollar
      - 6C   # Canadian Dollar / US Dollar
      - 6S   # Swiss Franc / US Dollar
    agriculture:
      - ZC   # Corn
      - ZS   # Soybeans
      - ZW   # Wheat (Chicago SRW)
      - ZM   # Soybean Meal
      - ZL   # Soybean Oil
    livestock:
      - LE   # Live Cattle
      - HE   # Lean Hogs
      - GF   # Feeder Cattle
  n_products: 30
  # Two sessions where one clearing venue's settlement file is absent. On each, the
  # missing products are missing at every tenor, and the two missing sets partition the
  # universe exactly: 2020-02-28 carries no CBOT, NYMEX or COMEX product (YM, the four
  # treasuries, five grains, four energy, four metals) and 2014-06-13 carries no CME
  # product (three equity index, six currencies, three livestock). Neighbouring sessions
  # trade normally, so this is a missing file rather than a thin market. Eleven of the
  # levels in stage 03 are carried as a percentile within the decision date, and a
  # percentile taken over one venue's products is not comparable with one taken over
  # thirty, so both dates are dropped rather than ranked against a partial universe.
  excluded_sessions:
    - 2020-02-28
    - 2014-06-13

decision:
  cadence: weekly_friday_close
  snapshot: settlement_price
  execution_delay: monday_open

# 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).
#
# ``initial_cash`` is sized so the full 30-product CME universe can fill
# positions under integer-share rules across the top_k_grid (5, 10) in
# long-short mode. The binding constraint is the high-notional equity-
# index e-minis: at top_k=10 long-short the per-position dollar budget
# is 5% × cash, and integer-rounding zeros out any product whose notional
# exceeds that budget. ES holdout notional ≈ $347k and NQ ≈ $513k drive
# the floor: $10M (per-position $500k) clears ES in the holdout and NQ
# in the validation window. NQ in the very-late holdout (Dec 2025,
# notional $513k) still integer-rounds out for that one product at
# top_k=10 LS — a residual integer-share footnote documented in the
# README "Margin model" section.
#
# Per-product initial+maintenance margin (futures_specs.yaml::
# {initial,maintenance}_margin_pct → ContractSpec.margin_pct) is wired
# through Engine.from_config(contract_specs=...); the engine auto-
# populates the broker's margin_pct_schedule so the margin draw scales
# with each bar's notional rather than the account-wide 50% default.
# allow_leverage is set in config/backtest/base.yaml.
#
# ``allocator_lookback`` is the bars-of-underlying-price window applied
# uniformly to every moment-based allocator (inverse_vol, risk_parity, hrp,
# mvo_ledoit_wolf). CME daily settlement bars → 63 ≈ 3 months. Keeps
# allocators comparable.
execution:
  initial_cash: 10_000_000       # sized for full 30-product universe at top_k_grid
  share_type: integer            # Futures contracts are integer
  allocator_lookback: 63         # 3 months of daily CME settlement bars

mapping:
  class: long_short_carry_rank
  position_state_space: long_short
  entry_logic: rank_by_carry_or_momentum
  sizing: equal_risk_or_notional

costs:
  class: material
  components: [commission, spread, roll_slippage]
  commission_per_contract: 2.0
  spread_ticks:
    liquid: 1
    illiquid: 2
  # Products charged the illiquid spread: live cattle, feeder cattle and soybean oil
  # quote wider than the rest of the universe. Everything else pays the liquid one tick.
  illiquid_products: [LE, GF, ZL]

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 others (equal_weight, score_weighted, inverse_vol) take the cheap path.
    # Ch16 signal-stage selection. Long-short equal-weight top-k on both
    # label horizons; long-short is enforced at the backtest config level
    # (allow_short_selling=true).
    # 30-product universe with long_short=True: top_k=20 selects 20 longs ∪
    # 20 shorts (40 positions) from 30 products → degenerate (collapses to
    # ~zero trades). Cap at 10 = floor(30/3) to keep selections non-degenerate.
    top_k_grid:
      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.
    # No equal_weight entry. Equal weight IS the Chapter 16 baseline, and `stage` is not
    # part of `backtest_hash`, so an equal-weight run at the allocation stage is a no-op
    # reweight that hashes identically to its baseline parent - whichever writes first
    # claims the row and the other is lost. Measured in this case study's own pre-rebuild
    # store: 48 rows carrying `stage='signal'` with `allocation.method='equal_weight'`,
    # and zero `allocation`/`equal_weight` rows, which downstream tooling then reported
    # as both missing and unexpected trials. The baseline is already in the pool that the
    # later stages select over.
    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; commission + exchange fees + slippage
    # combined).
    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 for 03_financial_features. Every window the notebook
# uses is declared here and read from here, and each family carries the ten fields
# the register renders: what it reads, how far back, with what delay, and what it
# is claimed to do. ``role`` separates a signal — something the carry thesis says
# ranks products against each other — from a state variable describing the
# environment a ranking is formed in. Windows are CME settlement sessions.
features:
  windows:
    momentum: [21, 63, 126, 252]     # each carries a risk-adjusted twin
    short_return: 5                  # the weekly return, no Sharpe twin
    volatility: [21, 63, 126]
    skip_recent: 21                  # skip-month momentum runs t-252 to t-21
    skip_start: 252
    carry_smoothing: 21
    carry_zscore: [63, 126]
    carry_momentum: [5, 21]
    trend_sign: [63, 126, 252]
    moving_average: [21, 50, 200]
    high: 252                        # dist_from_52w_high
    low: 126                         # dist_from_6m_low
    rsi: 14
    yang_zhang: 21
    variance_ratio: {horizon: 5, window: 63}
  # Source column -> the name its within-date percentile is written under. The
  # names predate this register and later stages select by them, so the mapping
  # is explicit rather than a suffix rule.
  ranked:
    ret_21d: mom_rank_21d
    ret_63d: mom_rank_63d
    ret_126d: mom_rank_126d
    ret_252d: mom_rank_252d
    sharpe_63d: sharpe_rank_63d
    sharpe_126d: sharpe_rank_126d
    sharpe_252d: sharpe_rank_252d
    vol_21d: vol_rank
    rsi_14: rsi_rank
    carry_pct: carry_rank
    curve_curvature_norm: curvature_rank
  composites:
    momentum_composite: [mom_rank_63d, mom_rank_126d, mom_rank_252d]
    sharpe_composite: [sharpe_rank_63d, sharpe_rank_126d, sharpe_rank_252d]
  thresholds:
    carry_regime_band: 0.01          # carry, on the x12 scale, separating the two regimes from flat
    signal_long: 60                  # percentile of the carry-momentum composite
    signal_short: 40
    roll_proximity_day: 25           # calendar day from which the roll week is flagged
  # Planting, harvest, heating and cooling cycles are documented for these two
  # sectors and not for the financial ones.
  seasonal_sectors: [agriculture, energy]
  families:
    - name: term structure
      pattern: carry_pct|carry_21d|carry_momentum_*|carry_zscore_*|carry_regime_num|curve_curvature_norm|curvature_21d
      role: signal
      hypothesis: A product whose curve is in backwardation earns the roll the shape implies
      inputs: unadjusted front, second and third settlement prices
      lookback: 146
      lag: 0
      frame: term structure
      representation: front-to-next spread on a x12 scale, its smoothed level, z-score and change
      failure_mode: the spread inverts around a supply shock faster than the smoothing follows
    - name: momentum
      pattern: ret_*d|skip_month_mom|mom_accel_*|ts_mom_*
      role: signal
      hypothesis: Trends in futures persist over horizons of a quarter to a year
      inputs: roll-adjusted settlement price
      lookback: 252
      lag: 0
      frame: time series
      representation: simple return, its sign, and differences between horizons
      failure_mode: reverses over the most recent month, which skip-month momentum drops
    - name: risk-adjusted momentum
      pattern: sharpe_[0-9]*d
      role: signal
      hypothesis: A trend earned with less dispersion repeats more reliably than a raw one
      inputs: log return
      lookback: 252
      lag: 0
      frame: time series
      representation: mean log return over its own dispersion, annualized
      failure_mode: unbounded as dispersion approaches zero
    - name: volatility and regime
      pattern: vol_[0-9]*d|vol_ratio_*|vol_yz_*|vr_*d
      role: state
      hypothesis: Dispersion and the trending-or-reverting regime set what a ranking can earn
      inputs: OHLC settlement bars, log return
      lookback: 126
      lag: 0
      frame: time series
      representation: annualized standard deviation, ratios of windows, variance ratio
      failure_mode: close-to-close misses the overnight gap Yang-Zhang is here to catch
    - name: trend and range
      pattern: rsi_[0-9]*|ma_ratio_*|dist_from_*
      role: signal
      hypothesis: Price against its own recent path separates a trend from an exhausted one
      inputs: roll-adjusted settlement price
      lookback: 252
      lag: 0
      frame: time series
      representation: bounded oscillator, ratio to a moving average or a rolling extreme
      failure_mode: saturates in a sustained trend, where every product reads the same
    - name: calendar and season
      pattern: month_*|day_of_year_norm|quarter|roll_proximity|is_seasonal_sector
      role: state
      hypothesis: Agricultural and energy curves move with planting, harvest and heating cycles
      inputs: the session date and the product's sector
      lookback: 1
      lag: 0
      frame: calendar
      representation: circular month encoding, quarter, roll-week and sector flags
      failure_mode: says when in the year it is and never which product to prefer
    - name: cross-sectional position
      pattern: '*_rank|mom_rank_*|sharpe_rank_*|carry_rank_sector'
      role: signal
      hypothesis: Only relative standing is tradable in a long-short cross-sectional strategy
      inputs: the levels this register's other families produce
      lookback: 252
      lag: 0
      frame: cross-section
      representation: percentile within the decision date and contract position
      failure_mode: discards the level the state families carry instead
    - name: composite
      pattern: momentum_composite|sharpe_composite|carry_mom_composite|carry_mom_interaction|risk_adj_score|ls_signal
      role: signal
      hypothesis: Carry and momentum are separate premia, so their agreement is itself information
      inputs: the percentiles of the cross-sectional family
      lookback: 252
      lag: 0
      frame: cross-section
      representation: mean of percentiles, their product, and a banded discretization
      failure_mode: hides which of the two components moved the score

# 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.
#
# A fitted feature is bounded by the schedule below and not by a walk-forward
# period. The parameters behind a value are estimated from sessions strictly
# before it, refreshed on the cadence given, and the same value comes out
# whichever period later selects the row - so the artifact carries no fold column.
#
# The spectral features in section C.2 have no entry here. A Fourier transform of
# a window estimates nothing, so it has no estimation window to declare; its
# lookback is a lookback like the ones under `features.windows`.
model_based:
  arima:
    # Sessions of a product's own carry z-score spent before the first forecast.
    # A trading year, so the first order is chosen from a full year of the series
    # rather than from a few weeks of it. The walk was already causal at this
    # cadence before the schedule was declared here; what moved is where the two
    # numbers live.
    burnin: 252
    # How often the weights are re-estimated. Monthly.
    refit_every: 21
    # (p, d, q), declared rather than selected per window.
    #
    # This replaced AutoARIMA's stepwise search, and the measurement that decided it is
    # worth stating because it is the reason rather than the justification. Sampling
    # eight refit cutoffs on each of the 30 eligible products, 240 searches in all,
    # AutoARIMA returned 34 distinct orders and NO product held a single order across
    # its own walk. The modal choice (2,0,1) took 11.7% of searches and (2,0,2) 10.4%;
    # the best any product managed was 6 of its 8 cutoffs on one order. An order that
    # changes almost every month on an expanding window of the same series is the
    # information criterion moving with the sample, not structure being found, and it
    # makes "the ARIMA carry forecast" a different quantity in every block.
    #
    # (2,0,1) is the modal selection and the more parsimonious of the two that are
    # indistinguishable from each other.
    #
    # d=0 is by construction, not by search: carry_zscore is already a rolling z-score
    # clipped to +/-5, so it is stationary before this model sees it. 83% of the searches
    # agreed (199 of 240 chose d=0); the 17% that differenced it were over-differencing a
    # bounded series, which is a known failure of automatic selection rather than a
    # property of carry.
    order: [2, 0, 1]
  hmm:
    # Backwardation and contango.
    n_states: 2
    # Sessions of the book-wide carry average 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 rather than once per product.
    #
    # The average itself begins 2011-04-27, after the carry smoothing and z-score
    # windows, and the oldest period trains from the first session of the panel -
    # so this burn-in comes out of that period's training window rather than
    # ahead of it. The first regime value lands 2013-04-09: 504 of the 1,978
    # sessions in that period's training window, and no evaluation session, since
    # the earliest evaluation window opens 2019-01-03. The walk makes 44
    # estimates before the holdout freezes it.
    burnin: 504
    # How often the chain is re-estimated. A quarter, as in etfs and fx_pairs:
    # regime parameters are the slowest-moving thing this notebook fits.
    refit_every: 63

evaluation:
  n_splits: 5
  train_size: 8Y
  val_size: 1Y
  holdout_start: '2024-01-01'
  holdout_end: '2025-12-31'
  calendar: CME
  periods_per_year: 252  # CME 5d/wk

labels:
  primary: fwd_ret_5d
  buffer: 5D
  variants:
    - fwd_ret_21d
  variant_buffers:
    fwd_ret_21d: 21D
  # Vectorized-backtest thinning step per label: number of schedule slots
  # to advance per trade so holding periods don't overlap. Authored from
  # (schedule cadence, label horizon); add an entry here for any new label.
  rebalance_step:
    fwd_ret_5d: 1   # weekly_friday schedule, 5d horizon <= 7d gap
    fwd_ret_21d: 3  # weekly_friday schedule, 21d horizon -> ceil(21/7) = 3

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
  latent_factors:
    persistent_entities: true
    # Declared, not inferred. Without this key case_study.py:93 falls back to
    # preferred_latent_device(), which is `cuda` where torch sees a GPU and `cpu` where it
    # does not. device enters computation.runtime and computation.numerical_runtime, both
    # inside the hashed computation, so the same declared configuration resolves a different
    # training identity on a GPU host than on a CPU one and each publishes under the same
    # population name. 10b is neural and takes this; 10a overrides it to cpu, because
    # PCAModel is numpy and scipy with no device argument and recording cuda there would
    # describe a computation that did not happen.
    device: cuda
    model_kwargs:
      sdf:
        checkpoint_epochs: [256, 512, 768, 1024]  # conditional-relative; publishes global 256..1280
        beta_checkpoint_epochs: [256]
        beta_default_checkpoint: 256

causal:
  treatment: carry_pct
  # Bars the treatment's own construction window spans, which is what the placebo block has
  # to cover: permuting carry_pct 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:
  #
  # `(c0_price - c1_price) / c0_price * 12` in 04_model_based_features, from two prices at
  # the same timestamp. Nothing rolls, so the construction window spans one bar, which is what
  # this field declares. It is not a claim that the column is serially independent: carry_pct is
  # highly persistent, and 12_model_analysis measures that profile. Where the construction window
  # is the shorter of the two scales, as it is here, the label buffer is what sizes the block -
  # `block_size = max(label_buffer, treatment_window)`.
  treatment_window: 1
  confounders: [vol_21d, momentum_composite, carry_rank]
  method: walk_forward_dml

```