math-backend / cvxpy_engine.py
engineportf's picture
Upload folder using huggingface_hub
558db1e verified
Raw
History Blame Contribute Delete
34.4 kB
import numpy as np
import pandas as pd
import cvxpy as cp
import warnings
from typing import Optional, List, Dict, Any, Tuple
from dataclasses import dataclass
from config import Color, logger
from core_types import PortfolioState, ForecastResult
@dataclass
class TaggedConstraint:
name: str
constraint: cp.Constraint
display: bool = True
@dataclass
class CVXPYResult:
weights: Optional[pd.Series]
display_constraints: Dict[str, float]
binding_constraints: Dict[str, float]
relaxation_log: List[str]
error: Optional[str] = None
class CVXPYOptimizationEngine:
def __init__(self,
tickers: List[str],
returns_df: pd.DataFrame,
forecast: ForecastResult,
state: PortfolioState,
cfg: dict,
macro: Optional[dict],
spread_map: Optional[dict],
risk_input: int,
risk_factor: float,
capital: float,
adv_proxy: float,
safe_min: float,
asset_max: float,
sector_limit: float,
allow_shorts: bool,
durations: np.ndarray,
b_min: float,
b_max: float,
has_basis: bool,
max_turnover: float,
stability_spreads: np.ndarray,
stab_lambda: float,
silent: bool = False,
opt_params: Any = None):
self.tickers = tickers
self.n = len(tickers)
self.returns_df = returns_df
self.forecast = forecast
self.state = state
self.cfg = cfg
self.macro = macro
self.spread_map = spread_map
self.risk_input = risk_input
self.risk_factor = risk_factor
self.capital = capital
self.adv_proxy = adv_proxy
self.safe_min = safe_min
self.asset_max = asset_max
self.sector_limit = sector_limit
self.allow_shorts = allow_shorts
self.durations = durations
self.b_min = b_min
self.b_max = b_max
self.has_basis = has_basis
self.max_turnover = max_turnover
self.stability_spreads = stability_spreads
self.stab_lambda = stab_lambda
self.silent = silent
self.opt_params = opt_params
self.max_assets = self.cfg.get("max_assets", 0)
self.dropped_indices = []
self.solver_feas_tol = 5e-4
# Prepare core matrices
self.Sigma = self._prepare_covariance()
self.exp_rets_arr = np.nan_to_num(self.forecast.expected_returns.values, nan=0.0, posinf=0.0, neginf=0.0)
self.ret_matrix = self._prepare_return_matrix()
self.vol_arr = np.nan_to_num(self.forecast.covariance_result.volatility.values, nan=0.0, posinf=0.0, neginf=0.0)
self.beta_vals = np.nan_to_num(self.forecast.betas.values if self.forecast.betas is not None else np.ones(self.n), nan=1.0)
# Base variables
self.w = cp.Variable(self.n)
self.dw = cp.Variable(self.n)
self.short_w = cp.Variable(self.n)
self.sell_w = cp.Variable(self.n)
self.t_impact = None
if isinstance(self.state.current_weights, pd.Series):
self.prev_w = self.state.current_weights.drop(labels=['CASH'], errors='ignore').reindex(self.tickers).fillna(0.0).values
elif self.state.current_weights is not None and len(self.state.current_weights) > 0:
self.prev_w = np.asarray(self.state.current_weights)[:self.n]
else:
self.prev_w = np.zeros(self.n)
def _prepare_covariance(self) -> np.ndarray:
Sigma_raw = self.forecast.covariance_result.covariance.values
Sigma_raw = np.nan_to_num(Sigma_raw, nan=0.0, posinf=0.0, neginf=0.0)
Sigma = (Sigma_raw + Sigma_raw.T) / 2.0
Sigma = Sigma + np.eye(self.n) * 1e-8
if getattr(self.opt_params, 'use_fast_ewm_cov', False):
return Sigma
eigvals_psd, eigvecs_psd = np.linalg.eigh(Sigma)
Sigma = (eigvecs_psd * np.maximum(eigvals_psd, 1e-8)) @ eigvecs_psd.T
Sigma = (Sigma + Sigma.T) / 2.0
return Sigma
def _prepare_return_matrix(self) -> np.ndarray:
ret_matrix = np.nan_to_num(self.returns_df.values, nan=0.0, posinf=0.0, neginf=0.0)
if hasattr(self.forecast, 'garch_info') and self.forecast.garch_info:
for i, ticker in enumerate(self.tickers):
multiplier = np.sqrt(self.forecast.garch_info.get(ticker, {}).get('scale', 1.0))
ret_matrix[:, i] *= multiplier
return ret_matrix
def _apply_warm_start(self):
e2e_ws = self.cfg.get('_e2e_warm_start')
if e2e_ws is not None:
try:
ws_arr = np.asarray(e2e_ws, dtype=float).flatten()[:self.n]
if ws_arr.shape == (self.n,) and np.all(np.isfinite(ws_arr)):
self.w.value = ws_arr
if not self.silent:
logger.info("E2E warm-start applied to CVXPY solver.")
except Exception as e:
logger.warning(f"E2E warm-start failed: {e}")
def _build_risk_measure(self):
risk_measure = self.cfg.get('_risk_measure', 'Mean-Variance')
T_obs_daily = len(self.returns_df)
baseline_rf = self.cfg.get('baseline_risk_factor', 3.0)
risk_tilt = baseline_rf / max(self.risk_factor, 0.5)
cvar_alpha = self.cfg.get('cvar_alpha', 0.95)
cvar_lambda = self.cfg.get('cvar_lambda', 0.5) * (0.2 + 0.8 * risk_tilt)
regime_severity = self.macro.get("hmm_regime", {}).get("severity_score", 1.0) if self.macro else 1.0
if regime_severity > 1.5:
cvar_alpha = 0.99
cvar_lambda *= regime_severity
if cvar_alpha >= 0.999:
cvar_alpha = 0.99
risk_penalty = 0.0
risk_measure_constraints = []
VOL_TARGETS = {
1: 0.03, 2: 0.05, 3: 0.07, 4: 0.09, 5: 0.12,
6: 0.15, 7: 0.18, 8: 0.21, 9: 0.24, 10: 0.28
}
if risk_measure == 'Mean-Variance':
target_vol = VOL_TARGETS.get(int(round(self.risk_factor)), 0.15)
vol_slack = cp.Variable(nonneg=True)
risk_penalty = 100.0 * vol_slack
portfolio_variance = cp.quad_form(self.w, cp.psd_wrap(self.Sigma))
risk_measure_constraints = [portfolio_variance <= (target_vol**2) + vol_slack]
elif risk_measure == 'CVaR':
VaR = cp.Variable(1)
u_var = cp.Variable(T_obs_daily)
risk_penalty = cvar_lambda * (VaR[0] + (1.0 / (T_obs_daily * (1.0 - cvar_alpha))) * cp.sum(u_var))
risk_measure_constraints = [u_var >= -(self.ret_matrix @ self.w) - VaR[0], u_var >= 0]
elif risk_measure == 'CDaR':
C_rets = self.ret_matrix @ self.w
C = cp.cumsum(C_rets)
M = cp.Variable(T_obs_daily)
cdar_cons = [M >= C, M[0] >= 0]
for i in range(1, T_obs_daily):
cdar_cons.append(M[i] >= M[i-1])
D = M - C
VaR_CDaR = cp.Variable(1)
u_cdar = cp.Variable(T_obs_daily)
risk_penalty = cvar_lambda * (VaR_CDaR[0] + (1.0 / (T_obs_daily * (1.0 - cvar_alpha))) * cp.sum(u_cdar))
risk_measure_constraints = cdar_cons + [u_cdar >= D - VaR_CDaR[0], u_cdar >= 0]
elif risk_measure == 'Max Loss':
max_loss_var = cp.Variable(1)
risk_penalty = self.risk_factor * max_loss_var[0]
risk_measure_constraints = [max_loss_var >= -(self.ret_matrix @ self.w)]
elif risk_measure == 'MAD':
mu_port = self.w.T @ self.exp_rets_arr
abs_dev = cp.Variable(T_obs_daily)
risk_penalty = self.risk_factor * cp.sum(abs_dev) / T_obs_daily
risk_measure_constraints = [abs_dev >= (self.ret_matrix @ self.w) - mu_port, abs_dev >= mu_port - (self.ret_matrix @ self.w)]
elif risk_measure == 'Semi-Variance':
downside_dev = cp.Variable(T_obs_daily)
risk_penalty = (self.risk_factor / 2) * cp.sum_squares(downside_dev) / T_obs_daily
risk_measure_constraints = [downside_dev >= -(self.ret_matrix @ self.w), downside_dev >= 0]
else:
portfolio_variance = cp.quad_form(self.w, cp.psd_wrap(self.Sigma))
risk_penalty = (self.risk_factor / 2) * portfolio_variance
# Conditional CVaR Constraint Formulation
use_cvar = self.cfg.get('cvar_enabled', True)
cvar_cost = 0.0
cvar_constraints = []
if use_cvar and risk_measure == 'Mean-Variance':
cvar_ret_matrix = self.ret_matrix
if regime_severity > 1.5 and self.macro:
hmm_details = self.macro.get("hmm_regime", {}).get("details", {})
state_seq = hmm_details.get("state_sequence")
if state_seq is not None:
crash_mask = np.array(state_seq) == 2
if len(crash_mask) == self.ret_matrix.shape[0]:
slice_mask = crash_mask
elif len(crash_mask) > self.ret_matrix.shape[0]:
slice_mask = crash_mask[-self.ret_matrix.shape[0]:]
else:
pad = np.zeros(self.ret_matrix.shape[0] - len(crash_mask), dtype=bool)
slice_mask = np.concatenate([pad, crash_mask])
n_crash = int(np.sum(slice_mask))
if n_crash >= 21:
cvar_ret_matrix = self.ret_matrix[slice_mask]
if not self.silent:
print(f" {Color.DIM}ℹ CVaR conditioned on {n_crash} crash-regime days (of {self.ret_matrix.shape[0]} total).{Color.RESET}")
T_cvar = cvar_ret_matrix.shape[0]
VaR = cp.Variable(1)
u = cp.Variable(T_cvar)
cvar_cost = cvar_lambda * (VaR[0] + (1.0 / (T_cvar * (1.0 - cvar_alpha))) * cp.sum(u))
cvar_constraints = [u >= -(cvar_ret_matrix @ self.w) - VaR[0], u >= 0]
return risk_penalty, risk_measure_constraints, cvar_cost, cvar_constraints
def _get_problem(self, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn, use_factors=True, active_cvar=None):
cons_dict = {}
def add_cons(name, cons, display=True):
cons_dict[name] = TaggedConstraint(name, cons, display)
add_cons("Fully Invested", (cp.sum(self.w) == 1.0), False)
base_gross_cap = self.cfg.get("gross_leverage_cap", 1.0)
adj_gross_cap = min(base_gross_cap * 1.5, base_gross_cap + (abs(cur_min) * 1.5)) if self.allow_shorts else base_gross_cap
add_cons("Gross Leverage Cap", (cp.norm(self.w, 1) <= adj_gross_cap), False)
add_cons(f"Single Asset Floor ({cur_min:.1%})", (self.w >= cur_min), False)
add_cons(f"Single Asset Cap ({cur_max:.1%})", (self.w <= cur_max), False)
if self.t_impact is not None:
impact_cons = [cp.norm(cp.vstack([2 * self.dw[i], self.t_impact[i] - 1]), 2) <= self.t_impact[i] + 1 for i in range(self.n)]
add_cons("Market Impact SOCP", impact_cons, False)
add_cons("Short Position Bounds", (self.short_w >= -self.w), False)
add_cons("DW Nonneg", (self.dw >= 0), False)
add_cons("Short W Nonneg", (self.short_w >= 0), False)
add_cons("Sell W Nonneg", (self.sell_w >= 0), False)
if self.dropped_indices:
add_cons(f"Cardinality Pruning", (self.w[self.dropped_indices] == 0), True)
if np.any(self.prev_w):
add_cons("Trade Size Tracking (Buys)", (self.dw >= self.w - self.prev_w), False)
add_cons("Trade Size Tracking (Sells)", (self.dw >= self.prev_w - self.w), False)
add_cons("Sell Volume Tracking", (self.sell_w >= self.prev_w - self.w), False)
if cur_turn < 10.0:
add_cons(f"Max Turnover ({cur_turn:.0%})", (cp.sum(self.dw) <= cur_turn), False)
else:
add_cons("Trade Size Tracking (Buys)", (self.dw >= self.w), False)
add_cons("Trade Size Tracking (Sells)", (self.dw >= -self.w), False)
add_cons("Sell Volume Tracking", (self.sell_w >= -self.w), False)
if cur_turn < 10.0:
add_cons(f"Max Turnover ({cur_turn:.0%})", (cp.sum(self.dw) <= cur_turn), False)
sector_map = self.cfg.get("sector_map", {})
for sector in set(sector_map.values()):
idx = [i for i, t in enumerate(self.tickers) if sector_map.get(t) == sector]
if idx:
if self.allow_shorts:
max_gross_allowed = max(0.80, abs(cur_min) * len(idx))
add_cons(f"Sector Gross Exposure ({sector})", (cp.norm(self.w[idx], 1) <= min(cur_sec * 2.0, max_gross_allowed)), True)
else:
add_cons(f"Sector Concentration Limit ({sector})", (cp.sum(self.w[idx]) <= cur_sec), True)
custom_constraints = self.cfg.get("custom_constraints") or []
for cc in custom_constraints:
asset = cc.get("asset")
limit = float(cc.get("limit", 0))
direction = cc.get("direction", "max")
if asset in self.tickers:
idx = self.tickers.index(asset)
if direction == "max":
add_cons(f"Custom Max ({asset} <= {limit:.1%})", (self.w[idx] <= limit), True)
elif direction == "min":
add_cons(f"Custom Min ({asset} >= {limit:.1%})", (self.w[idx] >= limit), True)
elif asset and asset.startswith("Sector:"):
sector = asset.split(":", 1)[1].strip()
idx = [i for i, t in enumerate(self.tickers) if sector_map.get(t) == sector]
if idx:
if direction == "max":
add_cons(f"Custom Sector Max ({sector} <= {limit:.1%})", (cp.sum(self.w[idx]) <= limit), True)
elif direction == "min":
add_cons(f"Custom Sector Min ({sector} >= {limit:.1%})", (cp.sum(self.w[idx]) >= limit), True)
if active_cvar is None:
active_cvar = []
ff_betas = self.forecast.ff_betas
if use_factors and self.cfg.get("factor_neutrality", False) and ff_betas is not None:
if 'HML' in ff_betas.columns:
hml_loadings = np.nan_to_num(ff_betas['HML'].values)
add_cons("Factor Neutrality (HML)", (cp.abs(self.w.T @ hml_loadings) <= 0.05), True)
if 'SMB' in ff_betas.columns:
smb_loadings = np.nan_to_num(ff_betas['SMB'].values)
add_cons("Factor Neutrality (SMB)", (cp.abs(self.w.T @ smb_loadings) <= 0.05), True)
if use_factors and self.cfg.get("risk_budgeting", False) and np.any(self.prev_w):
port_var_prev = self.prev_w.T @ self.Sigma @ self.prev_w
if port_var_prev > 1e-6:
marginal_risk = (self.Sigma @ self.prev_w) / port_var_prev
add_cons("Risk Budgeting (Max 15%)", (cp.multiply(self.w, marginal_risk) <= 0.15), True)
flat_cons = []
for tc in cons_dict.values():
if isinstance(tc.constraint, list):
flat_cons.extend(tc.constraint)
else:
flat_cons.append(tc.constraint)
cons_list = flat_cons + active_cvar + self.risk_measure_constraints
if use_beta:
eff_b_min = max(0.0, self.b_min - 0.3) if widen_beta else self.b_min
eff_b_max = self.b_max + 0.3 if widen_beta else self.b_max
beta_cons_labeled = {
f"Beta Target Min ({eff_b_min:.1f})": (self.w.T @ self.beta_vals >= eff_b_min),
f"Beta Target Max ({eff_b_max:.1f})": (self.w.T @ self.beta_vals <= eff_b_max)
}
for k, v in beta_cons_labeled.items():
add_cons(k, v, True)
cons_list += list(beta_cons_labeled.values())
if use_dur and any(self.durations > 0):
min_dur = self.cfg.get("min_duration", 0.0)
max_dur = self.cfg.get("max_duration", 50.0)
port_duration = self.w.T @ self.durations
if min_dur > 0:
add_cons(f"Min Duration ({min_dur}y)", (port_duration >= min_dur), True)
if max_dur < 50.0:
add_cons(f"Max Duration ({max_dur}y)", (port_duration <= max_dur), True)
cons_list += [cons_dict[k].constraint for k in cons_dict if "Duration" in k]
if active_cvar:
dynamic_obj = cp.Maximize(self.base_objective_term - self.cvar_cost)
else:
dynamic_obj = cp.Maximize(self.base_objective_term)
return cp.Problem(dynamic_obj, cons_list), cons_dict, cons_list
def _solution_violations(self, w_val, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn):
vals = np.asarray(w_val, dtype=float)
violations = []
if vals.shape != (self.n,) or not np.all(np.isfinite(vals)):
return ["solution contains non-finite or mis-shaped weights"]
total_weight = float(np.sum(vals))
gross_weight = float(np.sum(np.abs(vals)))
base_gross_cap = self.cfg.get("gross_leverage_cap", 1.0)
adj_gross_cap = min(base_gross_cap * 1.5, base_gross_cap + (abs(cur_min) * 1.5)) if self.allow_shorts else base_gross_cap
if total_weight > 1.0 + self.solver_feas_tol:
violations.append(f"sum(w)={total_weight:.4f} exceeds 100% budget")
if gross_weight > adj_gross_cap + self.solver_feas_tol:
violations.append(f"gross={gross_weight:.4f} exceeds cap {adj_gross_cap:.4f}")
if float(np.min(vals)) < cur_min - self.solver_feas_tol:
violations.append(f"min weight {float(np.min(vals)):.4f} below {cur_min:.4f}")
if float(np.max(vals)) > cur_max + self.solver_feas_tol:
violations.append(f"max weight {float(np.max(vals)):.4f} above {cur_max:.4f}")
if cur_turn < 10.0:
prev = self.prev_w if np.any(self.prev_w) else np.zeros(self.n)
turnover = float(np.sum(np.abs(vals - prev)))
if turnover > cur_turn + self.solver_feas_tol:
violations.append(f"turnover={turnover:.4f} exceeds cap {cur_turn:.4f}")
sector_map = self.cfg.get("sector_map", {})
for sector in set(sector_map.values()):
idx = [i for i, t in enumerate(self.tickers) if sector_map.get(t) == sector]
if not idx:
continue
if self.allow_shorts:
sector_gross = float(np.sum(np.abs(vals[idx])))
sector_gross_cap = min(cur_sec * 2.0, 0.80)
if sector_gross > sector_gross_cap + self.solver_feas_tol:
violations.append(f"{sector} sector gross={sector_gross:.4f} exceeds cap {sector_gross_cap:.4f}")
else:
sector_weight = float(np.sum(vals[idx]))
if sector_weight > cur_sec + self.solver_feas_tol:
violations.append(f"{sector} sector={sector_weight:.4f} exceeds cap {cur_sec:.4f}")
if use_beta:
eff_b_min = max(0.0, self.b_min - 0.3) if widen_beta else self.b_min
eff_b_max = self.b_max + 0.3 if widen_beta else self.b_max
beta = float(vals @ self.beta_vals)
if beta < eff_b_min - self.solver_feas_tol or beta > eff_b_max + self.solver_feas_tol:
violations.append(f"beta={beta:.4f} outside target {eff_b_min:.4f}-{eff_b_max:.4f}")
if use_dur and any(self.durations > 0):
port_duration = float(vals @ self.durations)
min_dur = self.cfg.get("min_duration", 0.0)
max_dur = self.cfg.get("max_duration", 50.0)
if min_dur > 0 and port_duration < min_dur - self.solver_feas_tol:
violations.append(f"duration={port_duration:.4f} below {min_dur:.4f}")
if max_dur < 50.0 and port_duration > max_dur + self.solver_feas_tol:
violations.append(f"duration={port_duration:.4f} above {max_dur:.4f}")
return violations
def _solve_prob(self, prob, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn, use_factors=True, is_deep_relaxation=False):
if np.any(self.prev_w):
self.w.value = self.prev_w
try:
solvers_to_try = [cp.CLARABEL] if is_deep_relaxation else [cp.CLARABEL, cp.HIGHS]
for solver in solvers_to_try:
if solver not in cp.installed_solvers():
continue
try:
solve_opts = {"warm_start": True}
if solver == cp.SCS:
solve_opts.update({"eps_abs": 1e-5, "eps_rel": 1e-5, "eps_gap": 1e-5})
if solver == cp.OSQP:
solve_opts.update({"max_iter": 100000})
if not self.silent:
logger.debug(f"Attempting solver {solver}...")
prob.solve(solver=solver, **solve_opts)
if not self.silent:
logger.debug(f"Solver {solver} finished with status {prob.status}.")
if self.w.value is not None:
logger.debug(f"Solver {solver} returned weights shape {np.shape(self.w.value)}")
if prob.status in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE]:
break
except Exception as exc:
if not self.silent:
logger.warning(f"Solver {solver} failed: {exc}")
continue
if prob.status not in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE]:
if not self.silent:
logger.warning(f"No viable solution found. Final status {prob.status}.")
return False, f"Solver status: {prob.status}"
if self.w.value is not None:
w_val = np.asarray(self.w.value).flatten()
w_val = np.nan_to_num(w_val, nan=0.0, posinf=0.0, neginf=0.0)
if w_val.shape != (self.n,) or not np.all(np.isfinite(w_val)):
return False, "Non-finite or mis-shaped weights"
self.w.value = w_val
violations = self._solution_violations(self.w.value, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn)
if not violations:
if prob.status == cp.OPTIMAL_INACCURATE and not self.silent:
logger.warning(f"Solver converged to OPTIMAL_INACCURATE. Constraint violations were within accepted tolerances ({self.solver_feas_tol}).")
return True, ""
old_tol = self.solver_feas_tol
self.solver_feas_tol = 1e-3
violations_1e3 = self._solution_violations(self.w.value, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn)
self.solver_feas_tol = old_tol
if not violations_1e3 and violations:
if not self.silent:
logger.warning(f"Solution had violations in the [5e-4, 1e-3] band and was rejected: {violations}")
else:
if not self.silent:
logger.warning(f"Solution rejected due to specific violations: {violations}")
return False, f"Constraint violations: {violations}"
return False, "Weights are None"
except Exception as exc:
logger.error(f"Optimization solver exception: {exc}")
raise
def solve(self) -> CVXPYResult:
self._apply_warm_start()
# Build base objective components
expected_portfolio_return = self.w.T @ self.exp_rets_arr
rfr = self.cfg.get("risk_free_rate", 0.0)
cash_yield = (1.0 - cp.sum(self.w)) * rfr
from execution import calibrate_gamma
ac_gamma = self.cfg.get("tc_volume_profile", None)
if ac_gamma is None:
# Dynamically calculate gamma for each asset
gammas = np.array([calibrate_gamma(self.adv_proxy, v) for v in self.vol_arr])
impact_coeffs = gammas * self.vol_arr * np.sqrt(self.capital / self.adv_proxy)
else:
impact_coeffs = ac_gamma * self.vol_arr * np.sqrt(self.capital / self.adv_proxy)
market_impact_penalty = 0.0
if ac_gamma is None or ac_gamma > 0:
self.t_impact = cp.Variable(self.n, nonneg=True)
market_impact_penalty = cp.sum(cp.multiply(impact_coeffs, self.t_impact))
spreads_arr = np.array([self.spread_map.get(t, 0.0008) for t in self.tickers]) if self.spread_map else np.full(self.n, 0.0008)
tc = self.cfg.get("transaction_cost", 0.001)
friction_arr = spreads_arr + tc
tc_lambda = 1.0 / max(self.risk_factor, 0.1)
tc_cost = tc_lambda * cp.sum(cp.multiply(self.dw, friction_arr))
borrow_cost = self.cfg.get('short_borrow_cost', 0.0)
borrow_drag = borrow_cost * cp.sum(self.short_w)
stab_penalty = self.stab_lambda * cp.sum(cp.multiply(self.stability_spreads, cp.square(self.w))) if self.stab_lambda > 0 else 0.0
tax_cost = 0.0
if self.has_basis:
tax_cost = cp.sum(cp.multiply(self.sell_w, self.state.gain_fractions * self.state.tax_rates))
risk_penalty, self.risk_measure_constraints, self.cvar_cost, cvar_constraints = self._build_risk_measure()
self.base_objective_term = (
expected_portfolio_return
+ cash_yield
- risk_penalty
- tc_cost
- market_impact_penalty
- tax_cost
- borrow_drag
- stab_penalty
)
use_cvar = self.cfg.get('cvar_enabled', True)
def run_relaxation_loop():
prob = None
relaxation_log = []
_ss = self.safe_min if not self.allow_shorts else -1.0
stage_params = [
("Base", True, False, True, True, self.safe_min, self.asset_max, self.sector_limit, self.max_turnover, None),
("Widen Beta", True, True, True, True, self.safe_min, self.asset_max, self.sector_limit, self.max_turnover, "Widened portfolio beta targets."),
("Drop Beta", False, False, True, True, self.safe_min, self.asset_max, self.sector_limit, self.max_turnover, "Dropped portfolio beta target constraints."),
("Drop Duration", False, False, False, True, self.safe_min, self.asset_max, self.sector_limit, self.max_turnover, "Dropped portfolio duration targets."),
("Drop Factors", False, False, False, False, self.safe_min, self.asset_max, self.sector_limit, self.max_turnover, "Dropped Factor Neutrality/Risk Budgeting."),
("Remove Min Weights", False, False, False, False, _ss, self.asset_max, self.sector_limit, self.max_turnover, "Removed minimum asset weight limits."),
("Relax Sector Caps", False, False, False, False, _ss, self.asset_max, min(1.0,self.sector_limit*2), self.max_turnover, "Doubled sector concentration limits."),
("Remove Turnover", False, False, False, False, _ss, self.asset_max, min(1.0,self.sector_limit*2), 2.0, "Removed maximum turnover constraint."),
("Unconstrained", False, False, False, False, _ss, 1.0, 1.0, 2.0, "Removed maximum single-asset weight limits."),
]
for cvar_pass, cvar_active in enumerate([True, False]):
relaxation_log = []
active_cvar = list(cvar_constraints) if (use_cvar and cvar_active) else []
for stage_idx, (name, use_beta, widen_beta, use_dur, use_factors, cur_min, cur_max, cur_sec, cur_turn, log_msg) in enumerate(stage_params):
prob, cons_dict, cons_list = self._get_problem(use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn, use_factors, active_cvar)
is_deep = stage_idx >= 5
solved, reason = self._solve_prob(prob, use_beta, widen_beta, use_dur, cur_min, cur_max, cur_sec, cur_turn, use_factors, is_deep_relaxation=is_deep)
if solved:
if log_msg:
relaxation_log.append(log_msg)
if stage_idx >= 7:
print(f"\n {Color.YELLOW}⚠ WARNING: Optimization dropped into 'Unconstrained' mode. Original constraints abandoned due to mathematical infeasibility.{Color.RESET}")
if not self.silent:
logger.warning("Optimization dropped into 'Unconstrained' mode.")
elif stage_idx >= 4:
print(f"\n {Color.YELLOW}⚠ WARNING: Optimization hit deep relaxation (Stage {stage_idx}). Constraints heavily modified.{Color.RESET}")
if not self.silent:
logger.warning(f"Optimization hit deep relaxation (Stage {stage_idx}).")
return prob, cons_dict, cons_list, relaxation_log
relaxation_log.append(f"Stage '{name}' failed: {reason}")
if log_msg:
relaxation_log.append(log_msg)
if prob is not None and not self.silent:
logger.debug(f"All relaxations failed. Final prob.status was: {prob.status}")
return None, None, None, relaxation_log
if not self.silent:
print(f" Solving convex optimization (β range {self.b_min:.1f}{self.b_max:.1f})...", end="", flush=True)
prob, final_cons_dict, final_cons_list, relaxation_log = run_relaxation_loop()
# --- Cardinality Constraint Heuristic ---
if prob is not None and prob.status in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE] and self.max_assets > 0:
w_val = np.asarray(self.w.value).flatten()
active_mask = np.abs(w_val) > 1e-4
if np.sum(active_mask) > self.max_assets:
num_to_drop = int(np.sum(active_mask)) - self.max_assets
active_indices = np.where(active_mask)[0]
sorted_by_weight = active_indices[np.argsort(np.abs(w_val[active_indices]))]
drop_candidates = list(sorted_by_weight[:num_to_drop])
feasible_drop = False
while len(drop_candidates) > 0 and not feasible_drop:
self.dropped_indices = drop_candidates
prob_card, cons_dict_card, cons_list_card, log_card = run_relaxation_loop()
if prob_card is not None and prob_card.status in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE]:
prob = prob_card
final_cons_dict = cons_dict_card
final_cons_list = cons_list_card
relaxation_log.extend(log_card)
relaxation_log.append(f"Applied Cardinality Constraint (max {self.max_assets} assets), pruned {len(drop_candidates)} assets.")
feasible_drop = True
else:
drop_candidates.pop()
if not feasible_drop:
relaxation_log.append("Failed to safely apply cardinality constraint (infeasible bounds). Maintaining continuous weights.")
self.dropped_indices = []
prob, final_cons_dict, final_cons_list, log_card = run_relaxation_loop()
# ---------------------------------------
if prob is None:
return CVXPYResult(None, {}, {}, relaxation_log, "All relaxation stages exhausted.")
if prob.status not in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE] or self.w.value is None:
return CVXPYResult(None, {}, {}, relaxation_log, "Solver failed to find optimal status.")
best_w = pd.Series(self.w.value.copy(), index=self.tickers)
display_constraints: Dict[str, float] = {}
binding_constraints: Dict[str, float] = {}
if prob.status in [cp.OPTIMAL, cp.OPTIMAL_INACCURATE] and len(final_cons_list) > 0:
display_constraints = self._diagnose_binding_constraints(final_cons_dict)
binding_constraints = display_constraints
return CVXPYResult(best_w, display_constraints, binding_constraints, relaxation_log)
def _diagnose_binding_constraints(self, cons_dict, tolerance=1e-4):
binding = {}
if cons_dict is None:
return binding
for name, tc in cons_dict.items():
if not tc.display:
continue
c = tc.constraint
if isinstance(c, list):
duals = []
for sub_c in c:
if hasattr(sub_c, 'dual_value') and sub_c.dual_value is not None:
duals.extend(np.array(sub_c.dual_value).flatten())
if duals:
vals = np.array(duals)
sig_vals = vals[np.abs(vals) > tolerance]
if len(sig_vals) > 0:
binding[name] = float(np.mean(sig_vals))
continue
if hasattr(c, 'dual_value') and c.dual_value is not None:
vals = np.array(c.dual_value).flatten()
sig_vals = vals[np.abs(vals) > tolerance]
if len(sig_vals) > 0:
binding[name] = float(np.mean(sig_vals))
return binding