STRATA — Market State Framework

A trainable, stateful model architecture specialised for market trading — analogous to LSTM/GRU but built for regime estimation and decision support, not price prediction.

pip install strata-market

---

Quick Start — High Level (v2.4+)

from strata import StrataModel, StrataTrainer

# ── Use a pretrained model (zero setup) ──────────────────────────────────
model = StrataModel.from_pretrained("AAPL")   # also: TSLA, SPY, NVDA, QQQ
result = model.predict(candles)
# {"action": "LONG", "confidence": 0.71, "regime": "TRENDING",
#  "approved": True, "guard_reason": "", "state": {...}}

# ── Train your own model from OHLCV data ─────────────────────────────────
trainer = StrataTrainer(asset="MSFT")
windows = StrataTrainer.prepare_windows(my_candles)   # sliding windows
model   = trainer.train(windows, n_trials=100)
model.save("msft_model.json")

# ── Load and deploy anywhere ─────────────────────────────────────────────
model  = StrataModel.load("msft_model.json")
result = model.predict(latest_candles)
print(result["action"])   # "LONG" / "SHORT" / "HOLD"

---

Quick Start — Low Level

from strata import initial_state, update_state, decide, StrataGUARD, StrataMEMORY, sense

# 1. Initialize
state  = initial_state()
memory = StrataMEMORY()
guard  = StrataGUARD(asset="AAPL")   # optional: asset-aware guard thresholds

# 2. Per-bar update loop
candles = [...]  # list of OHLCV dicts, oldest → newest, minimum 3

inp        = sense(candles)           # raw OHLCV → semantic signals
mem_signal = memory.snapshot()        # pattern memory state
state      = update_state(state, inp, mem_signal)
decision   = decide(state)            # {"action": "LONG"|"SHORT"|"HOLD", "confidence": ..., "regime": ...}

approved, reason = guard.evaluate(state, decision, decision["confidence"])
final_action = decision["action"] if approved else "HOLD"

print(f"{final_action}  ({reason})")

---

StrataModel — Trainable Model

StrataModel.from_pretrained(ticker) — zero setup

model = StrataModel.from_pretrained("AAPL")  # AAPL | TSLA | SPY | NVDA | QQQ
print(model.summary())

StrataTrainer — train from your own data

from strata import StrataModel, StrataTrainer

# Your candles: flat list of OHLCV dicts, oldest → newest
candles = [{"open": ..., "high": ..., "low": ..., "close": ..., "volume": ...}, ...]

# Prepare sliding windows (window_size=31 recommended)
windows = StrataTrainer.prepare_windows(candles, window_size=31)

# Train
trainer = StrataTrainer(asset="MSFT", verbose=True)
model   = trainer.train(windows, n_trials=100)

# Save — share with anyone who has strata-market installed
model.save("msft_model.json")

StrataModel.load() — deploy anywhere

model  = StrataModel.load("msft_model.json")
model.reset()   # start fresh session

# Per-bar prediction
result = model.predict(candles_window)
# {
#   "action":       "LONG",      # LONG | SHORT | HOLD
#   "confidence":   0.71,
#   "regime":       "TRENDING",  # TRENDING | RANGING | TRANSITIONING | CHOPPY
#   "risk":         "LOW",       # LOW | MEDIUM | HIGH
#   "approved":     True,        # False = GUARD blocked the action
#   "guard_reason": "",          # e.g. "TRAP_RISK_HIGH (0.67 > 0.60 [RANGING])"
#   "state":        {...},       # full internal state snapshot
# }

# Feed outcome to circuit breaker
model.record_outcome(was_loss=False)

---

StrataNet — PyTorch Neural Network (v2.5+)

STRATA-NET is a novel recurrent neural network architecture purpose-built for market trading — not a vanilla LSTM or GRU, but a new architecture where the hidden state directly encodes interpretable market concepts.

Hidden State Design

h = [bias, momentum, trap_risk, uncertainty]   ← 4 dimensions, always interpretable

bias        ∈ [-1, 1]   directional conviction (negative = bearish, positive = bullish)
momentum    ∈ [ 0, 1]   structural energy from breakouts
trap_risk   ∈ [ 0, 1]   adverse selection / liquidity trap pressure
uncertainty ∈ [ 0, 1]   volatility-driven ambiguity

Unlike LSTM/GRU hidden states (opaque vectors), these dimensions have enforced hard bounds and semantic meaning at every step of inference.

Architecture

Input: OHLCV window  (B, T, 5)
         │
         ▼
StrataEmbedding      learned OHLCV → semantic features  (replaces hand-crafted SENSE)
         │
         ▼
StrataCoreCell  ×T   custom GRU-like cell with 4 structured gates
         │             — update gate z, reset gate r (like GRU)
         │             — but projected back to 4-dim interpretable hidden state
         │             — bounds enforced per dimension every step
         ▼
StrataHead           hidden state → LONG/SHORT/HOLD + confidence + regime
         │
         ▼
Output: action logits (3) + confidence (1) + regime logits (4)

Quick Start

import torch
from strata import StrataNet, StrataNetConfig, StrataNetTrainer, StrataNetDataset

# ── Train from OHLCV candles (no manual labeling needed) ─────────────────
# Labels are auto-generated by the STRATA state machine (teacher-student)
dataset = StrataNetDataset.from_candles(my_candles, seq_len=30, asset="AAPL")
trainer = StrataNetTrainer(asset="AAPL")
model   = trainer.train(dataset, epochs=30)

# Save as .pt — standard PyTorch format
model.save("aapl_net.pt")

# ── Load and deploy ───────────────────────────────────────────────────────
model = StrataNet.load("aapl_net.pt")

# Predict from a (30, 5) tensor of normalised OHLCV features
result = model.predict_action(x_tensor)
# {
#   "action":     "LONG",          # LONG | SHORT | HOLD
#   "confidence": 0.71,
#   "regime":     "TRENDING",      # TRENDING | RANGING | TRANSITIONING | CHOPPY
#   "state": {
#       "bias":        0.89,       # interpretable — always in [-1, 1]
#       "momentum":    0.43,       # interpretable — always in [0, 1]
#       "trap_risk":   0.21,       # interpretable — always in [0, 1]
#       "uncertainty": 0.54,       # interpretable — always in [0, 1]
#   }
# }

Normalise your OHLCV before feeding to StrataNet

from strata.net_trainer import _normalise_window

window  = candles[-30:]   # last 30 candles
x       = torch.tensor(_normalise_window(window), dtype=torch.float32)  # (30, 5)
result  = model.predict_action(x)

Custom config

cfg = StrataNetConfig(
    asset        = "TSLA",
    seq_len      = 60,      # longer lookback for volatile assets
    embed_dim    = 64,      # larger embedding for more complex patterns
    core_expand  = 32,      # wider recurrent cell
    head_dim     = 32,
    dropout      = 0.15,
)
model = StrataNet(cfg)
print(model.summary())
# StrataNet v1.0
#   asset       : TSLA
#   seq_len     : 60
#   hidden_dim  : 4  [bias, momentum, trap_risk, uncertainty]
#   parameters  : ~18,000

Comparison: STRATA vs LSTM/GRU/Transformer

	LSTM / GRU	Transformer	StrataNet
Domain	General	General	Market trading only
Hidden state	Opaque N-dim vector	Attention weights	4-dim, fully interpretable
State bounds	None	None	Hard bounds per dimension
Output	Price / probability	Token / probability	LONG/SHORT/HOLD + regime
Label source	Future price needed	Manual labels	Auto from STRATA teacher
Model file	.pt	.pt / .bin	.pt
Parameters	Millions	Millions	~5,000 (default)

---

Architecture (State Machine — Classic API)

Raw OHLCV
    │
    ▼
┌─────────┐
│  SENSE  │  OHLCV → {trend, vol, liquidity_above, break_structure}
└────┬────┘
     │
     ▼
┌─────────┐     ┌──────────┐
│  CORE   │◄────│  MEMORY  │  pattern bank (Short/Mid/Long layers)
└────┬────┘     └──────────┘
     │
     ▼
┌─────────┐
│ DECIDE  │  state → {action, confidence, risk, regime}
└────┬────┘
     │
     ▼
┌─────────┐
│  GUARD  │  hard-rule override (NOT learned — explicit, auditable)
└────┬────┘
     │
     ▼
  Final Action
     │
     ▼
┌─────────┐
│  LOOP   │  P&L feedback → weight adaptation (optional)
└─────────┘

---

Modules

strata.sense(candles) → signals

Converts a window of OHLCV candles into semantic signals.

Signal	Range	Meaning
trend	[-1, 1]	Directional EMA crossover, ATR-normalized
vol	[0, 1]	Volatility intensity (ATR / median price)
liquidity_above	[0, 1]	Volume spike vs rolling average
break_structure	[0, 1]	Price closing beyond prior N-bar high/low

from strata import sense

candles = [
    {"open": 150.0, "high": 151.2, "low": 149.5, "close": 150.8, "volume": 1_200_000},
    # ... at least 3 candles, oldest → newest
]
signals = sense(candles)
# {"trend": 0.12, "vol": 0.03, "liquidity_above": 0.15, "break_structure": 0.0}

---

strata.update_state(state, inp, memory) → state

Core state transition. Updates the internal state vector based on current signals and memory.

from strata import initial_state, update_state

state = initial_state()   # {"bias": 0, "momentum": 0, "trap_risk": 0, "uncertainty": 0}
state = update_state(state, signals, memory_snapshot)
# {"bias": 0.08, "momentum": 0.04, "trap_risk": 0.12, "uncertainty": 0.02, "regime_score": 0.06}

State dimensions:

Key	Range	Meaning
bias	[-1, 1]	Directional conviction. Positive = bullish, negative = bearish
momentum	[0, 1]	Structural energy from breakouts and volume
trap_risk	[0, 1]	Probability of adverse selection / liquidity trap
uncertainty	[0, 1]	Volatility-driven ambiguity in current conditions

---

strata.decide(state) → decision

Produces a structured action proposal from current state.

from strata import decide

decision = decide(state)
# {"action": "LONG", "confidence": 0.62, "risk": "LOW", "regime": "TRENDING"}

Regimes: TRENDING | RANGING | TRANSITIONING | CHOPPY

---

strata.StrataGUARD

Hard-rule risk gate. Blocks actions that violate explicit risk constraints. Rules are regime-aware and asset-aware.

from strata import StrataGUARD

guard = StrataGUARD(asset="TSLA")   # HIGH vol: relaxed min_abs_bias
guard = StrataGUARD(asset="SPY")    # LOW vol ETF: relaxed max_trap_risk
guard = StrataGUARD()               # no asset profile — use regime defaults

approved, reason = guard.evaluate(state, decision, decision["confidence"])
# (True,  "OK")
# (False, "TRAP_RISK_HIGH (0.67 > 0.60 [RANGING])")

guard.record_outcome(was_loss=True)   # drives circuit breaker

Guard rules (checked in order):

Rule	Fires when
TRAP_RISK_HIGH	trap_risk > max_trap_risk
LOW_CONFIDENCE	confidence < min_confidence
CHAOS_REGIME	uncertainty > max_uncertainty
NO_CONVICTION	|bias| < min_abs_bias
CIRCUIT_BREAKER	consecutive_losses >= max

---

strata.StrataMEMORY

Three-layer pattern memory (Short → Mid → Long). Patterns are promoted by consistency and frequency.

from strata import StrataMEMORY

memory = StrataMEMORY()

# Feed observations each bar
memory.observe("fake_breakout", score=0.75)
memory.observe("trend_strength", score=0.60)
memory.step()   # advance time: promote, decay

# Use in state update
mem_signal = memory.snapshot()   # {"fake_breakout": 0.71, "trend_strength": 0.58}
state = update_state(state, signals, mem_signal)

---

strata.StrataLOOP

Closed-loop weight adaptation based on P&L feedback. Optional — STRATA works without it.

from strata import StrataLOOP

loop = StrataLOOP()

# After each closed trade:
loop.record_outcome(state, decision["action"], pnl=+0.012)

# Get adapted weights for next update
weights = loop.weights
state = update_state(state, signals, mem_signal, weights=weights)

print(loop)
# STRATA-LOOP | updates=47 | recent_wr=61.70% | step=0.00476 | outcomes=47

---

Supported Assets

STRATA ships with pre-calibrated volatility profiles for common assets:

Class	Assets	Adjustment
HIGH vol	TSLA, NVDA, AMD, COIN, PLTR, MSTR	min_abs_bias relaxed by -0.010
MED vol	AAPL, MSFT, GOOGL, META, AMZN, NFLX	No adjustment (baseline)
LOW vol ETF	SPY, QQQ, IWM, DIA, VTI, GLD, TLT	max_trap_risk relaxed by +0.10

Pass asset="TICKER" to StrataGUARD to activate profile. Unknown tickers default to MED.

---

Installation

# No external dependencies beyond Python 3.9+
# Clone and run directly:
git clone <repo>
cd STRATA-FRAMEWORK-MARKET-TRADING
py -3.11 tests/test_harness.py   # verify installation

---

Running Tests

# Core behavior (1000-step simulation)
py -3.11 tests/test_harness.py

# STRATA-LOOP weight adaptation
py -3.11 tests/test_loop.py

# Extreme scenario stress test (5 scenarios)
py -3.11 tests/test_stress.py

---

Running Calibration

# Single asset calibration (requires parquet data in data/raw/)
py -3.11 replay/multi_year_calibration.py --ticker AAPL --save

# Cross-asset comparison report (requires all 5 calibrations)
py -3.11 replay/comparison_report.py

# Walk-forward IS/OOS test
py -3.11 replay/walk_forward_test.py --ticker AAPL

---

Calibration Results (V2 Weights, 5-year data 2020–2024)

Asset	Guard Rate	HOLD%	LONG%	SHORT%	Guard Spread
AAPL	40.6%	53.0%	34.5%	12.5%	—
TSLA	37.8%	52.9%	36.1%	11.0%	—
SPY	39.9%	50.6%	36.2%	13.2%	—
NVDA	36.2%	50.6%	37.2%	12.1%	—
QQQ	39.3%	49.7%	37.4%	12.8%	—
Max-min spread					4.3% ✓

All assets: state drift STABLE, long memory formed, cross-year std < 2%.

---

Stress Test Results

Scenario	Guard Block	LONG%	Result
Crash	100%	0%	✓ SANE
Fake Breakout	11.5%	38%	✓ SANE
Sideways Chop	100%	0%	✓ SANE
Strong Trend	9.2%	90%	✓ SANE
Slow Grind	0.3%	99%	✓ SANE

---

Design Philosophy

STRATA is a cognitive layer, not a trading system. It answers one question: "what kind of market are we in right now, and what constraints should that impose on action?"

Key principles:

• State, not prediction — STRATA estimates regime, not price
• Hard rules, not learned risk — guard thresholds are explicit and auditable
• Structural adaptation, not per-asset retraining — one weight set, asset profiles for adjustment
• Bounded learning — LOOP weights stay within ±0.3 of calibrated seed values

See MANIFESTO.md for the full academic/theoretical treatment.

---

Project Structure

strata/
├── __init__.py          public API
├── sense.py             OHLCV → semantic signals
├── core.py              state transition engine + weights
├── memory.py            3-layer pattern memory
├── decide.py            state → action proposal
├── guard.py             hard-rule risk gate (regime + asset aware)
└── loop.py              closed-loop weight adaptation

replay/
├── multi_year_calibration.py    5-year calibration runner
├── comparison_report.py         V1 vs V2 + cross-asset report
└── walk_forward_test.py         IS/OOS generalization test

tests/
├── test_harness.py      core behavior checks (1000 steps)
├── test_loop.py         STRATA-LOOP unit tests
└── test_stress.py       5 extreme scenario tests

data/
├── raw/                 parquet OHLCV files (not included)
└── results/             calibration CSV outputs