AI for Raw Material Price Prediction: Planning Purchases 90 Days Ahead
A metal stamping company in Tennessee was spending $4.2 million per year on steel coil. Their purchasing process was straightforward: when inventory dropped to the reorder point, they called three suppliers and bought at the best current price. In the 18 months after implementing an AI-based price prediction system, they saved $186,000, about 4.4% of their annual steel spend, by timing purchases to avoid price peaks and locking in contracts during predicted troughs.
That 4.4% doesn't sound dramatic until you realize it came from a $35,000 software investment with zero changes to their suppliers, specifications, or manufacturing process.
What Raw Material Price Prediction Actually Does
The AI doesn't try to predict the exact price of hot-rolled coil steel on June 15th. It generates probability distributions over future price ranges: there's a 70% chance the price will be between $820 and $890 per ton in 60 days, a 15% chance it will be above $890, and a 15% chance it will be below $820. This probabilistic approach is more useful than a point estimate because it enables risk-adjusted purchasing decisions.
The models ingest multiple categories of data. Commodity futures prices and forward curves provide the market's current expectation of future prices. Supply-side indicators include raw material production data (iron ore output, aluminum smelter capacity utilization), logistics data (shipping rates, port congestion), and inventory levels at mills and service centers. Demand-side indicators include manufacturing PMI data, construction spending, automotive production forecasts, and order backlog data from major consuming industries.
Macroeconomic factors like currency exchange rates, interest rates, and trade policy changes also feed the model. Steel tariff announcements, for example, create predictable price spikes that a model trained on historical tariff impacts can anticipate.
Model Architecture and Accuracy
Most production systems use ensemble models combining several techniques. Gradient-boosted trees (XGBoost, LightGBM) handle the structured tabular data effectively. LSTM or transformer-based time series models capture temporal patterns and trends. Some systems include a natural language processing component that monitors news feeds for events likely to affect prices (trade disputes, natural disasters, plant closures).
At the 30-day horizon, well-calibrated models predict the price direction (up or down relative to current price) correctly 68% to 75% of the time. At 60 days, directional accuracy drops to 62% to 70%. At 90 days, it's 58% to 65%. These numbers may seem modest, but they're substantially better than naive forecasting (assuming prices stay the same, which is correct about 50% of the time for monthly changes) and significantly better than human judgment, which tends to be biased by recency and anchoring effects.
The magnitude prediction is less accurate than direction. The model might correctly predict that steel prices will increase in 60 days, but overestimate or underestimate the magnitude by 5% to 15%. For purchasing decisions, the direction is more important than the exact magnitude: knowing that prices are likely to increase justifies buying now rather than waiting.
Turning Predictions Into Purchasing Strategy
The practical output of the system is a set of recommendations for the purchasing team. When the model predicts rising prices with high confidence, it recommends accelerating purchases and potentially locking in forward contracts at current prices. When it predicts falling prices, it recommends buying minimum quantities and waiting for better pricing.
A manufacturing company using this approach needs to balance the price prediction against carrying costs. Buying an extra month's inventory of steel coil to lock in a lower price only makes sense if the expected price increase exceeds the carrying cost (warehouse space, insurance, capital cost of tied-up inventory, and the risk of material becoming obsolete if specifications change). The AI system incorporates these carrying costs into its recommendations.
For materials with liquid futures markets (steel, aluminum, copper, resins), the system can also recommend hedging strategies using financial instruments. A manufacturer who will need 200 tons of aluminum in 90 days can hedge against price increases by buying aluminum futures, and the AI's price prediction informs whether the hedging premium is worth paying.
Limitations and Honest Assessment
Raw material price prediction has real limitations that any honest assessment should acknowledge. Black swan events (pandemics, wars, major trade policy changes) can invalidate predictions entirely. The model's training data doesn't include scenarios it hasn't seen before, and the most expensive price movements are often driven by unprecedented events.
Market structure changes also challenge the models. The shift in steel pricing from quarterly contracts to spot-market-based pricing over the past decade means that models trained on older data may not capture current price dynamics. Regular model retraining (monthly or quarterly) is essential to maintain accuracy.
The data quality of input signals varies. Government economic statistics are released with delays (PMI data is monthly, industrial production data is monthly with revisions). By the time the data is published, the market has often already reacted to the underlying conditions. Real-time proxies (satellite data of shipping activity, power consumption at industrial facilities, web scraping of prices from supplier websites) can fill some of these gaps, but add complexity and cost.
For small and mid-size manufacturers spending less than $500,000 per year on any single material, the savings from AI-based price prediction may not justify the implementation cost. The sweet spot is companies with $2 million or more in annual spend on commodities with meaningful price volatility. At that scale, even a 3% to 5% improvement in purchasing timing generates returns that clearly exceed the system's cost.