Steel Industry Stochastic Procurement Optimization

Scenario-driven procurement and capacity planning for steel manufacturing — find the optimal balance between commitment and flexibility under genuine market uncertainty.

   

---

🎯 The Problem

Steel manufacturers commit to production capacity and raw material contracts months before knowing what demand, prices, or costs will be. Traditional ERP planning relies on point forecasts and safety stock buffers — ignoring correlations, failing to quantify risk, and leaving money on the table in every market regime.

This framework replaces that with a scenario-aware optimizer that evaluates thousands of plausible futures and finds the plan that maximizes expected profit while explicitly protecting against downside risk.

---

❓ What Questions It Answers

"How much base capacity should I commit to — and when should I use overtime?"

_{Optimal monthly split between base capacity and flexible production (overtime, subcontracting) across thousands of demand scenarios.}

"How should I balance fixed contracts, option call-offs, and spot purchases?"

_{Cost-optimal allocation across all three procurement instruments — committed volume, reserved-but-optional volume, and market-rate fallback — for the current price environment.}

"What is the real cost of playing it safe?"

_{Quantified profit–risk trade-off: see exactly how much expected profit you give up for each increment of downside protection.}

"How would this plan have held up in 2008… or COVID… or the 2022 price spike?"

_{Rolling backtest across 17 windows spanning 16 years — every crisis, every recovery, every supply shock from 2007 to 2023.}

"What if scrap spikes 30% and demand drops 20% at the same time?"

_{Analyst-defined stress scenarios are injected directly into the optimization — the plan is guaranteed to have seen them.}

"Is this actually better than what we do today?"

_{Measured head-to-head: the Value of Stochastic Solution (VSS) quantifies the profit improvement over traditional safety-stock planning in every regime tested.}

---

⚙️ How It Works

graph LR
    A["📊 <b>FRED Market Data</b><br/>20+ years of steel prices,<br/>scrap costs & demand"] --> B["🔧 <b>DataLoader</b><br/>Clean, align,<br/>log-returns"]
    B --> C["📈 <b>Regime-Switching VAR</b><br/>Fit normal + stress<br/>regime dynamics"]
    C --> D["🎲 <b>Scenario Generation</b><br/>3 000 correlated<br/>forward paths"]
    D --> E["✂️ <b>Scenario Reduction</b><br/>K-medoids → 300<br/>+ stress tails"]
    E --> F["⚙️ <b>Stochastic Optimizer</b><br/>Two-stage program<br/>with CVaR"]
    F --> G["📋 <b>Optimal Plan</b><br/>Capacity, contracts,<br/>risk metrics"]

    style A fill:#4472C4,stroke:#2F5597,color:#fff
    style B fill:#5B9BD5,stroke:#4472C4,color:#fff
    style C fill:#70AD47,stroke:#548235,color:#fff
    style D fill:#70AD47,stroke:#548235,color:#fff
    style E fill:#FFC000,stroke:#BF9000,color:#000
    style F fill:#ED7D31,stroke:#C55A11,color:#fff
    style G fill:#7030A0,stroke:#4E2074,color:#fff

Loading

Stage 1 (commit now): Base production capacity · Fixed procurement contracts · Framework option reservations
Stage 2 (adapt later): Framework exercise · Spot purchases · Flex production · Inventory management

---

📊 Results at a Glance

Backtested across 17 rolling windows from 2007 to 2023 — including the GFC crash, Euro debt crisis, COVID-19 demand shock, and post-COVID supply spike:

Metric	Stochastic Optimizer	Safety Stock Benchmark
Total profit (16-year backtest)	+€182M above benchmark	Baseline (€3,135M)
Profit uplift	+5.8% (Risk Neutral, λ=0.0)	—
Planning windows outperformed	14 / 17 (82%, Risk Neutral)	—
Fill rate	≥ 97%	85–95%
Worst-case profit (CVaR 5%)	Materially better	Fragile in crises
Shortage events	Near-zero	Common in volatile periods

Risk profile comparison (all vs. Safety Stock benchmark):

Profile	λ	VSS	Uplift	Windows won
Risk Neutral	0.0	+€182M	+5.8%	14/17
Moderate	0.3	+€156M	+5.0%	13/17
Conservative	0.7	+€163M	+5.2%	13/17
Full CVaR	1.0	+€168M	+5.3%	12/17

Risk Neutral wins both the most windows (14/17) and the highest total profit in this backtest — expected-profit maximization dominates when the scenario model captures regime structure well. Higher λ profiles trade some total upside for reduced tail exposure; Full CVaR wins fewer windows but outperforms by larger margins during volatile regimes (2008, COVID, post-COVID spike). The λ parameter is a strategic choice, not a calibration: the right answer depends on your market view and balance sheet tolerance for tail risk.

---

🔧 Key Features

Risk-Aversion Control A single parameter shifts the objective from pure expected-profit maximization to explicit downside protection:   Maximize expected profit, no tail-risk penalty   Hedge against worst outcomes while preserving upside   Prioritize protecting worst-case profit via CVaR	Stress Testing Inject analyst-defined extreme scenarios directly into the optimization so the plan is guaranteed to have seen them: "What if selling prices drop to the 5th percentile while scrap costs spike to the 95th?" Covers price collapses, cost spikes, demand crashes — and any combination of the three.
Three Contract Instruments Fixed contracts — committed volume at known price Framework call-offs — right but not obligation, bounded exercise price Spot purchases — unlimited, market-rate fallback	16-Year Rolling Backtest 17 autonomous replanning windows across real market history: GFC Crash · Euro Debt Crisis · Post-GFC Stagnation · Recovery · COVID-19 · Post-COVID Supply Spike Every regime tests a different failure mode for traditional planning.

---

👤 Who Is This For

• Supply chain leaders evaluating optimization-based planning vs. traditional ERP/MRP approaches
• Procurement managers looking to quantify the value of contract flexibility under price uncertainty
• Operations research practitioners seeking a production-ready two-stage stochastic programming implementation with real data
• Data scientists & quants interested in VAR-based scenario generation with stress testing for commodity markets

---

🚀 Quick Start

pip install -r requirements.txt

from src.data.loader import DataLoader
from src.scenario.regime_switching import RegimeSwitchingGenerator
from src.optimization.stochastic import StochasticOptimizationModel, ModelParameters
from src.params.risk import RiskProfile

# 1. Load 15 years of FRED market data
loader = DataLoader(fred_api_key="your_key")
loader.load_from_fred().subset(180, last_observation="2019-12-01")
loader.convert_to_real_prices("2019-12-01", {"P": 520, "C": 130, "D": 50000})
loader.compute_log_returns()

# 2. Fit Regime-Switching VAR → generate 3,000 scenarios → reduce to 300
gen = RegimeSwitchingGenerator(n_regimes=2)  # normal + stress regimes
gen.fit(loader)
print(gen.regime_summary())                  # inspect regime structure
reduced = gen.reduce(gen.generate(n_scenarios=3000, horizon=12), n_clusters=300)

# 3. Solve
result = StochasticOptimizationModel().run(
    reduced.scenarios, reduced.probabilities,
    params=ModelParameters(c_fix=320, x_fix_max=8000, x_opt_max=4000, alpha=1.2, pen_short=500),
    risk_profile=RiskProfile(risk_aversion=0.3),
    variable_mapping={"C": "c_spot", "P": "p_sell"},
)

print(f"E[Profit]  EUR {result.risk_metrics['Expected_Profit']:>12,.0f}")
print(f"CVaR (5%)  EUR {result.risk_metrics['CVaR_95']:>12,.0f}")
print(f"Sharpe         {result.risk_metrics['Sharpe']:>12.3f}")

Installation notes

• A free FRED API key is required for data loading.
• Windows: scikit-learn-extra (K-medoids) needs C++ build tools — install Microsoft C++ Build Tools or conda install -c conda-forge scikit-learn-extra.

---

⚠️ Limitations

• Two stages, not rolling — Stage 2 assumes full uncertainty resolution at once. Real planning involves continuous rebalancing with partial information. Multi-stage extensions would improve realism at significant computational cost.
• Single product — The current model handles one aggregate steel product. Multi-product extensions (e.g., flat vs. long products) are not yet supported.
• No supply disruption modeling — Scenarios capture price and demand uncertainty but not discrete supply chain disruptions (plant outages, port closures, trade sanctions).
• Solver dependency — Requires a linear programming solver. HiGHS (open-source, default) is sufficient; commercial solvers (Gurobi, CPLEX) are supported but not required.
• Monthly granularity only — the data pipeline, scenario generator, and optimization model all operate at monthly frequency. Sub-monthly or weekly planning horizons would require a redesigned data layer.

---

📖 Documentation

Document	Contents
Problem Formulation	Business context, contract types, production structure, why two-stage
Mathematical Formulation	VAR model, scenario generation, stochastic program, CVaR, benchmark math
Package Documentation	Full API reference with usage examples for every class and method

---

📊 Data Sources

All inputs are public data from the Federal Reserve Economic Data (FRED) API:

Variable	Series	Description
Steel price	WPU101704	PPI: Hot Rolled Bars, Plates & Structural Shapes
Scrap cost	WPU1012	PPI: Iron and Steel Scrap
Demand	IPG3311A2S	Industrial Production: Steel Products

---

MIT License — see LICENSE

Birge & Louveaux (2011) · Lütkepohl (2005) · Rockafellar & Uryasev (2000)