CPZ Python SDK

Strategy Framework, Backtesting Engine, Risk Guards, Execution Algorithms, and Multi-Broker Trading

What's New in v4.7.0 — Institutional Trust Suite

Five coordinated controls land together, each aligned to SEC Rule 15c3-5 (pre-trade) and SEC 17a-4 / FINRA 4511 (records) principles. This is a controls framework, not a claim of regulatory certification.

• Bounded Autonomy (client.mandates) — trading mandates that bound order flow per scope: symbol universe, per-order notional and quantity caps, gross exposure, leverage, daily-loss limit, and a mandatory expiry for live scopes. Enforced at every first-party order and deploy surface, backed by a hash-chained local audit ledger (cpz.audit) and mandate / killswitch / audit CLI subcommands.
• Backtest Trust Layer (client.backtests) — every backtest can emit a run card and validation artifact: metrics, benchmark context, data provenance, a deterministic config hash, and point-in-time / OHLC-sanity / lookahead-bias checks. The "SDK-attested" verdict is recomputed server-side, so a client can never self-certify.
• Factor Library (client.factors) — fetch the platform factor catalog and bench any factor over a universe and period (IC, rank IC, IR, turnover, with alive / reversed / dead status). Formulas run through an AST-whitelisted evaluator, never eval.
• Strategy Health (client.health) — a falsifiable-hypothesis registry plus a decay lifecycle (active / monitoring / decayed / disabled) computed from rolling Sharpe and IC. It records only; it never stops or modifies a deployment.
• Shadow Account (client.shadow) — parse broker execution history (Alpaca, IBKR Flex, generic CSV) into normalized fills, compute behavioral diagnostics (FIFO round trips, win rate, holding periods, disposition effect), and replay an honestly-bounded counterfactual.

Behavioral changes — read before upgrading:

• Live orders now require CPZ platform connectivity. The kill switch and mandate checks fail closed: if the control plane is unreachable, a live order raises rather than proceeding unchecked. Paper orders warn and proceed. Risk-reducing closes are exempt by design.
• validate=True is the new backtest default. Data-quality checks and the lookahead gate now run unless you pass validate=False.
• Removed the unauthenticated hardcoded-IP engine REST fallback, which cannot coexist with mandate enforcement.

Recent releases

• v4.6.6 — Bloomberg Data License market-data provider (client.data.bloomberg); order-fill durability.
• v4.6.3 — real, vendor-agnostic options chain with greeks (options() / get_option_chain()), routed through the CPZ backend instead of hitting Alpaca directly.
• v4.6.0 — implied-volatility history and IV rank (client.data.get_iv_history / iv_rank), vendor-agnostic and capability-based.
• v4.2.0 — client.data.scout(source, **params): one-call execution of any Data Scout starter snippet, credentials resolved server-side from your stored API connections.
• v4.1.0 — institutional FIX 4.4 client (cpz.fix) for direct broker connectivity, sync and async.
• v4.25.0: research Compute Fabric client and truthful local QAOA simulation; earlier real-QPU claims corrected.
• v3.0.0 — Strategy Framework, Backtest Engine, Pre-Trade Risk Guard, TWAP/VWAP/Iceberg execution algos.

See CHANGELOG.md for the full reconciled history (v2.4.0 → v4.7.0).

Overview

The CPZ Python SDK is the unified interface for the CPZ quantitative trading platform. Write strategies once and run them in backtest or live mode with zero code changes.

Installation

pip install cpz-ai                    # core SDK (includes classical quantum-inspired optimizers)
pip install cpz-ai[risk]              # + statsmodels for advanced risk analytics
pip install cpz-ai[quantum]           # + dwave-neal for simulated annealing
pip install cpz-ai[quantum-braket]    # + Amazon Braket local QAOA simulation (CPU)
pip install cpz-ai[all]               # + all optional dependencies

Quantum Portfolio Optimization

QUBO formulations support classical exact and annealing solvers, plus an experimental Braket local CPU simulator. Remote managed simulators and physical QPUs are disabled in cpz-quant 1.1.0. Earlier hardware and cost-tracking claims were unsupported by the implementation and are superseded here.

from cpz.portfolio import BraketSolver, qubo_portfolio_selection

# No AWS credentials or provider charges. Local CPU resources still have cost.
solver = BraketSolver(device="local", shots=256, max_iter=8, seed=7)
result = qubo_portfolio_selection(returns, target_k=2, solver=solver)

This does not demonstrate quantum advantage. QHRP permutation problems have N squared binary variables; local QAOA is bounded to 20 variables.

Compute Fabric research API (v4.25.0)

client.compute provides advisory recommendations, explicit small CPU hedge solves, batch planning, and benchmark uploads. client.quantum is an alias for the research interface, not physical QPU execution.

# Existing authenticated CPZClient. All numbers below are synthetic.
proposal = client.compute.optimize_hedge(
    backend="edge_wasm", algorithm="brute_force", request_id="hedge-demo-1",
    book_greeks={"DEMO": {"delta": 1, "gamma": 0, "vega": 0}},
    candidates=[{
        "symbol": "DEMO-HEDGE", "underlying": "DEMO", "lot": 1,
        "greeks_per_lot": {"delta": -1, "gamma": 0, "vega": 0},
        "notional_per_lot": 100, "spread_bps": 1,
    }],
    cardinality=1, penalty=10,
    greek_weights={"delta": 1, "gamma": 1, "vega": 1},
)
assert proposal.production_eligible is False

The endpoint requires platform schema activation, entitlement and a spending policy. It fails closed when these are unavailable. Keep the same request ID after an uncertain response; use a new ID only for intentional new work. Run uploads are caller-reported evidence, never billing authorization. Reported provider costs are not verified invoices.

Proposals validate linear Greeks only and cannot emit orders. Live orders declaring a research compute_run_id are blocked even in shadow mode. Untagged user-created orders use existing controls; manually copied numbers cannot be traced. GPU execution and physical-QPU dispatch remain unavailable.

Strategy Framework

Write your strategy once as a Strategy subclass. The same code runs against the backtest engine or live broker with zero changes.

Define a Strategy

from cpz import Strategy, StrategyConfig, StrategyRunner, CPZClient

class MomentumConfig(StrategyConfig):
    fast_period: int = 10
    slow_period: int = 30
    trade_size: float = 100

class MomentumStrategy(Strategy):
    def on_start(self):
        self.subscribe_bars(self.config.instruments[0], "1D")

    def on_bar(self, bar):
        prices = self.cache.bars(bar.symbol)
        if len(prices) < self.config.slow_period:
            return
        fast_ma = sum(b.close for b in prices[-self.config.fast_period:]) / self.config.fast_period
        slow_ma = sum(b.close for b in prices[-self.config.slow_period:]) / self.config.slow_period

        if fast_ma > slow_ma and not self.portfolio.is_long(bar.symbol):
            self.buy(bar.symbol, qty=self.config.trade_size)
        elif fast_ma < slow_ma and self.portfolio.is_long(bar.symbol):
            self.flatten(bar.symbol)

    def on_order_filled(self, event):
        self.log.info(f"Filled {event.symbol} @ {event.fill_price}")

    def on_position_closed(self, event):
        self.log.info(f"Closed {event.symbol}, realized P&L: ${event.realized_pnl:,.2f}")

Backtest

client = CPZClient()
runner = StrategyRunner(client, mode="backtest")
runner.add_strategy(MomentumStrategy(MomentumConfig(instruments=["AAPL"])))

result = runner.backtest(
    start="2024-01-01",
    end="2024-12-31",
    initial_cash=100_000,
    commission_pct=0.0003,    # 3 bps
    slippage_pct=0.0001,      # 1 bp
)
print(result.summary())

Output:

============================================================
  BACKTEST RESULTS
============================================================
  Period:            2024-01-01 → 2024-12-31
  Initial Cash:      $100,000.00
  Final Equity:      $112,450.00
  Total Return:      +12.45%
------------------------------------------------------------
  Sharpe Ratio:      1.842
  Sortino Ratio:     2.315
  Max Drawdown:      -4.23%
------------------------------------------------------------
  Total Trades:      47
  Win Rate:          63.8%
  Profit Factor:     2.14
============================================================

Backtest Realism (v4.29.0)

Every realism model is opt-in. With none set, a backtest behaves as before, and result.modeling says what that leaves out.

from cpz.strategy.sim import CashAccount, MarginAccount, VolumeParticipation, ibkr_fixed

result = runner.backtest(
    start="2024-01-01",
    end="2024-12-31",
    order_model="intrabar",                # limit, stop, stop_limit, trailing_stop, MOO, MOC
    fee_model=ibkr_fixed(),                # or alpaca_zero_commission(), tradestation_equities(), PerShareFee(...)
    fill_model=VolumeParticipation(0.05),  # at most 5% of each bar's volume; remainders carry forward
    account=CashAccount(),                 # or MarginAccount(debit_interest_rate=...)
)
print(result.modeling.text())

Orders (order_model="intrabar"). Bars only give open, high, low and close, so each rule takes the reading least favourable to the strategy:

An order placed while handling a bar can first fill on a bar that starts after that bar ends. Time in force: DAY, GTC (with optional expire_at), IOC, FOK. Expired, cancelled and rejected orders are listed on the result with their reasons.

Fees. PercentFee (the historic default), PerShareFee(rate, min, max_pct), SecSection31Fee and FinraTafFee (dated rate tables for sells, sourced from sec.gov and finra.org), plus presets ibkr_fixed(), alpaca_zero_commission() and tradestation_equities(tier). Rates carry their source URL and effective date in cpz/strategy/sim/fees.py. A fill dated before the first published rate raises; a fill after the last one uses the rate still in effect, and result.modeling records rates_verified_as_of (2026-09-17) with a note when any fill is later than that. Per-order minimums and caps are charged once per order across partial fills, as IBKR bills them. result.fee_totals breaks fees down by component.

Accounts. CashAccount buys only with settled cash (T+2 before 2024-05-28, T+1 after) and rejects short sales. MarginAccount applies Reg T 50% initial margin, FINRA Rule 4210 maintenance, a liquidation event on a maintenance breach, and daily debit interest at the rate you pass. An order over buying power is rejected with a reason, never resized.

Scheduled events. Register in on_start; on_timer(event) fires in backtests on the NYSE calendar (2010 to 2028, holidays and 13:00 early closes):

self.schedule.every("tick", minutes=15)
self.schedule.at_market_open("open30", offset_minutes=30)
self.schedule.at_market_close("rebalance", offset_minutes=10, frequency="month_end")  # daily, weekly, month_start, month_end

Consolidation. self.consolidate("AAPL", "15Min", handler) rolls 1-minute bars into 15-minute bars and only ever hands the handler closed bars.

Corporate actions. price_mode="raw", corporate_actions=corporate_action_source("massive") replays unadjusted Massive bars and applies splits (position rescaled, open orders cancelled) and cash dividends (receivable on the ex-date, cash on the pay date). Providers without split and dividend data raise. The default replays bars as the data path returns them (the CPZ gateway requests split-adjusted bars from Alpaca and Polygon).

See examples/backtest_realism.py.

Walk-Forward Optimization (v4.31.0)

The optimization is a call in your own script, so whoever reads the script sees what was searched and how:

result = runner.walk_forward(
    MyStrategy,
    param_grid={"lookback": [10, 20, 40], "z_entry": [1.5, 2.0]},
    config=MyConfig(instruments=["SPY"]),
    start="2019-01-01",
    end="2025-12-31",
    folds=5, train_bars=504, test_bars=126, anchored=False,
    objective="sharpe",   # "sortino", "calmar", "total_return" or a callable(BacktestResult) -> float
    warmup_bars=40,       # optional indicator history before each test window
)
print(result.summary())

For each fold, every combination is backtested on the train window and the best in-sample objective is chosen (ties go to the earliest combination). Only then does that combination run on the next test_bars bars. Test windows follow one another without overlap and never influence a choice. The folds end at the last bar of the data. anchored=True keeps the first train bar fixed and grows the train window by test_bars each fold.

Honesty rules.

• No lookahead. Selection sees only the fold's train bars. Warm-up bars come from before the test window; the strategy sees them, but every order placed on them is refused (result.warmup_refused_orders) and they are not scored.
• Every evaluated combination is a trial. result.trials is combinations x folds, and it deflates the Sharpe of the stitched out-of-sample returns (result.deflation: deflated Sharpe, expected maximum Sharpe under the null, PSR, DSR probability). The formula is the platform's schema-2 rigor block, so SDK and Builder runs are comparable.
• Everything is reported. result.folds (train and test dates, chosen params, in-sample objective, out-of-sample metrics, and both backtests), result.grid / grid_frame() (every combination's in-sample score on every fold), result.oos (the stitched out-of-sample BacktestResult), result.stability (how often the choice changed, and the objective at the chosen point's neighbouring grid points) and result.degradation (in-sample vs out-of-sample objective).
• Hard budget. More than 256 combinations raises WalkForwardBudgetError; a grid is never sampled or truncated. Too few bars for train_bars + folds * test_bars raises BacktestDataError with the numbers.
• Deterministic. Runs are serial; the same inputs give the same result.

Every realism option (order_model, fee_model, fill_model, account, calendar, price_mode, corporate_actions) reaches every inner backtest, with stateful models copied per run, and result.modeling.walk_forward records the method, folds, train, test and warm-up bars, objective, grid and trials. Bars are fetched once for the whole window.

Each fold's test run starts flat with initial_cash; out-of-sample returns are compounded across folds, and a position still open at a test window's end is marked to market at its last close with no exit costs, as at the end of any backtest.

upload=True (or client.backtests.upload(result)) uploads the stitched out-of-sample run with a schema-2 rigor block: deflatedSharpe.numTrials is the full trial count, walkForward.refit is true, and walkForward.anchored.oosSharpe / oosReturnPct carry the stitched out-of-sample figures (the platform reads them under that key for rolling walk-forwards too; walkForward.method says which ran).

See examples/walk_forward.py.

Live Trading (Same Code!)

runner = StrategyRunner(client, mode="live")
runner.add_strategy(MomentumStrategy(MomentumConfig(instruments=["AAPL"])))
runner.run()  # Blocks until Ctrl+C

Strategy Lifecycle Hooks

Strategy Actions

self.buy("AAPL", qty=100)                    # Market buy
self.sell("AAPL", qty=50)                     # Market sell
self.buy("AAPL", qty=10, order_type="limit", limit_price=150.00)
self.sell("AAPL", qty=10, order_type="trailing_stop", trail_percent=5, time_in_force="GTC")
self.buy("AAPL", qty=10, order_type="market_on_close")  # backtest: order_model="intrabar"
self.flatten("AAPL")                          # Close position
self.flatten_all()                            # Close all positions
self.cancel_all_orders()                      # Cancel open orders
self.emit_signal("crossover", value=1.0)      # Emit custom signal
self.subscribe_bars("AAPL", "1Min")           # Subscribe to bars
self.subscribe_quotes("AAPL")                 # Subscribe to quotes

Strategy Context

Every strategy has access to:

self.cache.bars("AAPL")           # Cached bar history
self.cache.latest_bar("AAPL")     # Most recent bar
self.portfolio.equity()           # Current equity
self.portfolio.cash               # Available cash
self.portfolio.is_long("AAPL")    # Position check
self.portfolio.positions()        # All positions
self.clock.utc_now()              # Current time (real or simulated)
self.log.info("message")          # Structured logging
self.config.instruments           # Strategy config

Pre-Trade Risk Guard

Intercepts every order before it reaches the broker. Any violation raises RiskGuardViolation and the order is rejected.

from cpz import CPZClient

client = CPZClient()

# Attach risk guard to execution
client.execution.set_risk_guard(
    max_order_value=50_000,         # Reject orders > $50K
    max_position_pct=0.10,          # Max 10% of equity per position
    max_daily_loss=-5_000,          # Halt trading if daily P&L < -$5K
    max_open_orders=20,             # Max concurrent open orders
    blocked_symbols=["GME", "AMC"], # Blacklist
    cooldown_seconds=5.0,           # Min 5s between orders per symbol
    max_orders_per_minute=30,       # Rate limit
)

# Orders are now validated before submission
order = client.execution.order(symbol="AAPL", qty=10, side="buy", strategy_id="my-strat")

Or use directly in a strategy:

from cpz import RiskGuard

class SafeStrategy(Strategy):
    def on_start(self):
        guard = RiskGuard(max_order_value=25_000, max_daily_loss=-2_000)
        self._router.set_risk_guard(guard)
        self.subscribe_bars("AAPL", "1D")

Execution Algorithms

Split large orders into smaller slices for better execution.

from cpz.execution.algos import TWAPAlgorithm, VWAPAlgorithm, IcebergAlgorithm

# TWAP — equal slices over time
algo = TWAPAlgorithm()
slices = algo.generate_slices(order_request, {
    "duration_minutes": 30,
    "num_slices": 10,
})

# VWAP — volume-weighted slicing with intraday profile
algo = VWAPAlgorithm()
slices = algo.generate_slices(order_request, {
    "duration_minutes": 60,
    "num_slices": 13,
})

# Iceberg — hidden quantity
algo = IcebergAlgorithm()
slices = algo.generate_slices(order_request, {
    "display_qty": 100,  # Only show 100 shares at a time
})

Typed Domain Model

Precision-safe value types that reject NaN, Infinity, and invalid data at construction time.

from cpz import Price, Quantity

# Validated, immutable
price = Price("150.25")                # From string (exact)
price = Price.from_float(150.25, precision=2)  # From float
qty = Quantity.from_int(100)

# Type-safe arithmetic
total = price * qty                    # Returns Money("15025.00", "USD")
spread = Price("150.50") - Price("150.25")  # Returns Price("0.25")

# Fail-fast on invalid data
Price(float("nan"))    # ValueError: Price must be finite
Quantity(-1)           # ValueError: Quantity cannot be negative

Event Bus

Thread-safe pub/sub with wildcard topic matching.

from cpz import EventBus

bus = EventBus()

# Subscribe with wildcards
bus.subscribe("bar.AAPL.*", handle_aapl_bars)     # Any timeframe for AAPL
bus.subscribe("order.**", handle_all_orders)       # All order events

# Publish
bus.publish("bar.AAPL.1Min", bar_event)

Quick Start

from cpz import CPZClient

client = CPZClient()

# ── Risk Analytics ──────────────────────────────────────
sharpe = client.risk.sharpe(daily_returns)
snapshot = client.risk.compute(daily_returns, spy_returns, weights)
mc = client.risk.monte_carlo(daily_returns, num_simulations=10000)

# ── Trading ─────────────────────────────────────────────
client.execution.use_broker("alpaca", environment="paper")
order = client.execution.order(symbol="AAPL", qty=10, side="buy", strategy_id="my-strat")

# ── Market Data ─────────────────────────────────────────
bars = client.data.bars("AAPL", timeframe="1D", limit=100)
quotes = client.data.quotes(["AAPL", "MSFT", "GOOGL"])
gdp = client.data.economic("GDP")

# ── Simons (Quant Strategist) ───────────────────────────
response = client.simons.chat("Analyze AAPL for momentum trading")
print(response.content)

Risk Analytics

Comprehensive quantitative risk computation powered by numpy/scipy. Install with pip install cpz-ai[risk].

On first use the risk-free rate is the newest 3-month Treasury yield (DGS3MO) from the CPZ us_macro warehouse. If that read fails, the SDK logs a risk_free_rate_unavailable warning and uses 0% until you call client.risk.set_risk_free_rate(...). All methods accept daily returns as a list of floats.

Core Metrics

from cpz import CPZClient

client = CPZClient()

# Individual metrics
sharpe = client.risk.sharpe(daily_returns)               # Annualized Sharpe ratio
sortino = client.risk.sortino(daily_returns)              # Sortino (downside vol only)
vol = client.risk.volatility(daily_returns)               # Annualized volatility (%)
mdd = client.risk.max_drawdown(daily_returns)             # Maximum drawdown (%)
b = client.risk.beta(daily_returns, spy_returns)          # Beta vs benchmark
a = client.risk.alpha(daily_returns, spy_returns)         # Jensen's alpha (annualized)

# Full risk snapshot (all metrics at once)
snapshot = client.risk.compute(
    daily_returns=portfolio_returns,
    benchmark_returns=spy_returns,
    position_weights={"AAPL": 0.4, "NVDA": 0.35, "BTC/USD": 0.25},
    total_exposure=100000,
)

Sharpe Inference (Lopez de Prado Framework)

Statistical significance testing for the Sharpe ratio. Implements all 5 corrections from Lopez de Prado (2012, 2018):

1. Non-normality — SE adjusted for skewness and kurtosis
2. Serial correlation — Lo (2002) autocorrelation adjustment
3. Probabilistic Sharpe Ratio (PSR) — P(true Sharpe > benchmark)
4. Deflated Sharpe Ratio (DSR) — Multiple-testing correction
5. Minimum Track Record Length — How long before the Sharpe is credible?

inference = client.risk.sharpe_inference(daily_returns, num_trials=20)
print(f"PSR (prob true SR > 0): {inference.psr:.1%}")
print(f"DSR (adjusted for 20 trials): {inference.deflated_sharpe}")
print(f"Min track record needed: {inference.min_track_record_months} months")

Value at Risk

var_95 = client.risk.parametric_var(daily_returns, confidence=0.95)
hist_var = client.risk.historical_var(daily_returns, confidence=0.95)
mc = client.risk.monte_carlo(daily_returns, num_simulations=10000, horizon_days=5)

Advanced Analytics

dd = client.risk.drawdown_analysis(daily_returns)        # Drawdown decomposition
tail = client.risk.tail_risk(daily_returns)              # Skewness, kurtosis, Cornish-Fisher VaR
sizing = client.risk.position_size(100000, returns)      # Kelly criterion + vol targeting
exposure = client.risk.factor_exposure(returns, factors) # OLS factor decomposition
impacts = client.risk.stress_test_all(weights)           # 7 historical crisis scenarios
rolling = client.risk.rolling_metrics(returns, spy, 20)  # Rolling Sharpe, vol, beta

Trading

Broker Configuration

from cpz import CPZClient

client = CPZClient()

client.execution.use_broker("alpaca", environment="paper")
client.execution.use_broker("alpaca", environment="live")
client.execution.use_broker("tradestation", environment="paper")
client.execution.use_broker("ibkr", environment="paper")
client.execution.use_broker("saxo", environment="live")
client.execution.use_broker("tastytrade", environment="live")
client.execution.use_broker("kalshi", environment="paper")  # Kalshi demo exchange
client.execution.use_broker("polymarket")
client.execution.use_broker("fidelity")  # SnapTrade brokers use the venue name

Supported broker names: alpaca, tradestation, ibkr, saxo, tastytrade, kalshi, polymarket, snaptrade and the SnapTrade venues (fidelity, schwab, robinhood, questrade, wealthsimple, td, etrade, webull, coinbase, vanguard, merrill, sofi, public, firstrade), plus hft_engine and cpz_proxy. TradeStation, Saxo and tastytrade keep their OAuth tokens server-side, and Kalshi its encrypted API key, so they always trade through the CPZAI execution gateway.

Kalshi symbols name one outcome of one market, "<MARKET_TICKER>:YES" or "<MARKET_TICKER>:NO", and limit_price is that outcome's price in dollars (0 to 1). A sell only reduces a held outcome; buy the other outcome to take that side. Kalshi has no market or stop orders: a market order needs limit_price and runs immediate-or-cancel. environment="paper" is Kalshi's demo exchange (demo API key), "live" is production.

client.execution.use_broker("kalshi", environment="paper")
client.execution.submit_order(
    symbol="KXFED-26OCT-T4.25:YES", qty=10, side="buy",
    order_type="limit", limit_price=0.56, time_in_force="gtc",
)

Order Placement

# Simple order
order = client.execution.order(
    symbol="AAPL",
    qty=10,
    side="buy",
    strategy_id="my-strategy"
)

# Full control
from cpz import OrderSubmitRequest, OrderSide, OrderType, TimeInForce

request = OrderSubmitRequest(
    symbol="AAPL",
    side=OrderSide.BUY,
    qty=10,
    order_type=OrderType.LIMIT,
    time_in_force=TimeInForce.GTC,
    limit_price=150.00,
    strategy_id="my-strategy"
)
order = client.execution.submit_order(request)

Account and Positions

account = client.execution.get_account()
print(f"Buying Power: ${account.buying_power:,.2f}")

positions = client.execution.get_positions()
for pos in positions:
    print(f"{pos.symbol}: {pos.qty} shares @ ${pos.avg_entry_price}")

Execution: one API for every broker

Every broker exposes the same methods, on client.execution (sync), AsyncCPZClient().execution and BrokerRouter. A broker that cannot do something raises BrokerOperationUnsupported (a NotImplementedError) before it touches your account: never None, never invented data, never a cancel followed by a failed replace. Ask first with capabilities().

from cpz import BrokerOperationUnsupported

client.execution.use_broker("saxo", environment="live")
caps = client.execution.capabilities()
if caps.close_position:
    client.execution.close_position("AAPL", percentage=50, strategy_id="my-strategy")

order = client.execution.order(
    symbol="AAPL", qty=10, side="sell", order_type="stop", stop_price=180.0,
    strategy_id="my-strategy",
)

result = client.execution.cancel_all_orders()
if not result.complete:
    print(result.failed, result.remaining)

try:
    client.execution.replace_order(order.id, OrderReplaceRequest(stop_price=182.0))
except BrokerOperationUnsupported as exc:
    print(exc.broker, exc.operation, exc.reason)

OrderSubmitRequest also takes trail_price / trail_percent (trailing stops, exactly one), extended_hours, venue_symbol_id (the broker's own instrument id: IBKR conid, Saxo Uic:AssetType, tastytrade Instrument Type|symbol, SnapTrade universal id, Polymarket token id, Kalshi TICKER:YES / TICKER:NO), position_effect (open / close) and asset_class. Stop orders need stop_price, stop-limit orders need stop_price and limit_price.

Capabilities through the execution gateway (GET /cpz/brokers is the source of truth; capabilities() reads it):

Native adapters (no CPZ_PROXY_ENABLED, or explicit credentials) describe themselves with their own, narrower capabilities(): the native IBKR adapter places market and limit orders only and has no order list, replace, closes or portfolio history; the native SnapTrade adapter is read-only; the native Polymarket adapter needs a limit_price on every order; the native Alpaca adapter does not stream quotes.

Execution Gateway Routing

With CPZ_PROXY_ENABLED=1 in the environment (hosted strategy runs set it), use_broker(<name>) for every gateway broker (alpaca, tradestation, ibkr, snaptrade and its venues, polymarket, saxo, tastytrade, kalshi) and the implicit Alpaca paper default behind client.execution.order(...) route through the CPZAI execution gateway (cpz/orders/execute) instead of talking to the broker from your process: broker credentials stay server-side, the pre-trade guard runs on every submission, and each order carries a client_order_id idempotency key. The broker name travels as the tag (broker=fidelity for a SnapTrade venue). Passing api_key_id, api_secret_key, oauth_token or access_token to use_broker keeps the native adapter, and hft_engine always keeps its own. router.routed_via tells you which path is active.

import os

os.environ["CPZ_PROXY_ENABLED"] = "1"
client.execution.use_broker("ibkr", environment="paper")
print(client.execution.router.routed_via)  # "cpz_proxy"

Every OrderSubmitRequest gets client_order_id = "cpz-<strategy8>-<uuid12>" unless you supply one (at most 64 characters from [A-Za-z0-9_.:-]). The returned Order carries it, plus idempotent_replay=True when the gateway answered from an earlier submission of the same id instead of placing a second broker order. Gateway outcomes are typed:

from cpz.execution.cpz_proxy import (
    CPZBrokerRejected,
    CPZOrderLaneBusy,
    CPZSubmitUnconfirmed,
)

try:
    order = client.execution.order(symbol="AAPL", qty=10, side="buy", strategy_id="my-strategy")
except CPZSubmitUnconfirmed as exc:
    # The order may be live under exc.client_order_id. Look it up with
    # get_order_by_client_id() or resend the identical request; never use a fresh id.
    ...
except CPZOrderLaneBusy as exc:
    # Another submission holds this symbol's lane and the SDK already retried
    # once. Retry the identical request after exc.retry_after_ms.
    ...
except CPZBrokerRejected as exc:
    # Definitive broker rejection: exc.code (for example wash_trade_rejected)
    # and exc.broker_code carry the reason. Not retryable as-is.
    ...

A timeout or connection error mid-submit is not treated as "not placed": the SDK looks the order up by client_order_id (twice, 500 ms apart) and returns it if the broker id is bound, and raises CPZSubmitUnconfirmed otherwise.

Quantitative Libraries (v2.5.0+)

Local Indicators

25+ indicators computed locally (no API calls):

from cpz.indicators import sma, ema, rsi, macd, bollinger, atr, vwap

fast = ema(prices, period=10)
slow = ema(prices, period=30)
rsi_val = rsi(prices, period=14)
bb = bollinger(prices, period=20)   # .upper, .middle, .lower
macd_val = macd(prices)             # .macd, .signal, .histogram

Portfolio Optimization

12 optimizers including HRP, Black-Litterman, and Mean-CVaR:

from cpz.portfolio import mean_variance, risk_parity, hierarchical_risk_parity

weights = risk_parity(returns_df)
weights = hierarchical_risk_parity(returns_df)

Signal Construction

from cpz.signals import vol_target, kelly, regime_filter, max_positions

sized = vol_target(raw_signals, returns, target_vol=0.15)
filtered = regime_filter(signals, returns, window=60)
limited = max_positions(signals, max_n=10)

Alpha Research

from cpz.alpha import information_coefficient, signal_decay, quantile_returns

ic = information_coefficient(signals, forward_returns)
decay = signal_decay(signals, returns, max_lag=20)

HFT Engine

Deploy strategies to the Rust HFT engine for autonomous microsecond-latency execution.

status = client.engine.status()
client.engine.deploy(
    strategy_id="my-strategy",
    symbols=["AAPL", "NVDA"],
    broker="alpaca",
    environment="paper",
)
client.engine.stop("my-strategy")

Data

# Stock/crypto bars
bars = client.data.bars("AAPL", timeframe="1D", limit=100)

# Multi-symbol history for backtesting
df = client.data.history(["AAPL", "MSFT", "NVDA"], timeframe="1D", limit=252)

# Quotes, news, options, economic data, filings, sentiment
quotes = client.data.quotes(["AAPL", "MSFT"])
news = client.data.news("TSLA", limit=10)
gdp = client.data.economic("GDP")
filings = client.data.filings("AAPL", form="10-K")
sentiment = client.data.sentiment("GME")

# Massive (formerly Polygon): stocks, options, indices, futures, crypto, forex
bars  = client.data.massive.get_bars("AAPL", timeframe="1d")
chain = client.data.massive.get_options_chain("AAPL")
quote = client.data.massive.realtime.get_quote("AAPL")   # refuses DELAYED data

# 18 data providers: Massive, Alpaca, TwelveData, FRED, SEC EDGAR,
# Yahoo Finance, Databento, CoinGecko, Finnhub, Alpha Vantage, and more

Data Connections Guide — how to connect each provider, then fetch, pull and stream from it (key resolution, entitlements, streaming support, troubleshooting).

Simons — Quantitative Trading Strategist

response = client.simons.chat("Analyze AAPL for momentum trading")
print(response.content)

for chunk in client.simons.stream("Write a mean-reversion backtest"):
    if chunk.type == "text":
        print(chunk.content, end="", flush=True)

Architecture

CPZClient
├── Strategy Framework (NEW in v3.0.0)
│   ├── Strategy          Base class with 20+ lifecycle hooks
│   ├── StrategyConfig    Serializable configuration
│   ├── StrategyRunner    Live and backtest orchestration
│   ├── BacktestEngine    Historical replay with fill simulation
│   ├── RiskGuard         Pre-trade order validation (10 rules)
│   └── TWAP/VWAP/Iceberg Execution algorithms
│
├── Typed Domain Model (NEW in v3.0.0)
│   ├── Price             Decimal-backed, fail-fast on NaN/Inf
│   ├── Quantity           Non-negative, precision-safe
│   ├── Money              Currency-tracked arithmetic
│   └── Events            BarEvent, FillEvent, PositionEvent, ...
│
├── risk               Portfolio risk analytics (numpy/scipy)
│   ├── compute()      Full risk snapshot
│   ├── monte_carlo()  Monte Carlo VaR (Rust-accelerated)
│   ├── sharpe_inference() Lopez de Prado framework
│   ├── stress_test_all()  7 historical crisis scenarios
│   └── [20+ more methods]
│
├── execution          Multi-broker trading
│   ├── use_broker()   Configure broker (Alpaca, IBKR, SnapTrade, Polymarket)
│   ├── order()        Place orders with risk guard validation
│   ├── set_risk_guard() Attach pre-trade risk rules
│   └── get_positions() Current positions
│
├── data               Market and reference data (18 providers)
│   ├── bars()         OHLCV price data
│   ├── history()      Multi-symbol DataFrames
│   ├── economic()     31 US macro series (us_macro); other FRED ids need your own key
│   └── [indicators]   100+ technical indicators via TwelveData
│
├── indicators         25+ local indicators (no API calls)
├── signals            Position sizing and signal filters
├── portfolio          12 portfolio optimizers
├── alpha              IC analysis, signal decay, quantile returns
│
├── simons             Quantitative trading strategist AI
└── engine             Rust HFT engine for autonomous execution

Configuration

Get credentials at ai.cpz-lab.com/settings.

Testing

pytest --cov=cpz --cov-report=term-missing

Support

• Platform: ai.cpz-lab.com
• Repository: github.com/CPZ-Lab/cpz-py
• Email: contact@cpz-lab.com

_{Built by CPZ}