Machine learning models deployed in trading systems don't operate in a vacuum — they are trained on the strategies and market structures that active traders have used for decades. Understanding what those strategies actually are helps MLOps practitioners build better feature sets, more meaningful evaluation metrics, and more relevant monitoring alerts. This guide surveys five interconnected techniques that sit at the intermediate layer of active trading: between the simplicity of buy-and-hold and the opacity of high-frequency algorithms.
Commodity and futures markets have a term structure — different contracts expiring at different dates trade at different prices simultaneously. When the market is in backwardation in futures markets, near-dated contracts trade above those expiring further out. This inverted structure often signals tight near-term supply: the market is willing to pay a premium for immediate delivery. Roll yield — the profit or loss from rolling an expiring contract into the next one — turns positive in backwardation, which is one reason commodity-focused funds pay close attention to curve shape rather than just spot prices.
In options markets, the calendar-spread options strategy exploits differences in time value across expiration dates. A classic setup involves selling a near-term option while buying a longer-dated option at the same strike. Because options lose time value at an accelerating rate as expiration approaches, the short leg decays faster than the long leg, creating profit for the spread seller — provided the underlying price stays near the strike. The strategy profits from relatively quiet, range-bound conditions, which connects it naturally to the next approach.
Betting on the spread between two related stocks — pairs trading — is one of the most widely taught relative-value strategies. The logic is mean-reversion: two historically correlated securities diverge in price for a short period, and the spread eventually reverts. A typical pairs trade shorts the temporarily expensive security and buys the temporarily cheap one. Calendar spreads and pairs trading both share a mean-reversion assumption: in calendars, time value reverts toward zero; in pairs, price spread reverts toward a historical equilibrium. Both strategies are also neutral to the overall market direction, which is part of their appeal for portfolio construction.
When a security has established a clear price floor and ceiling over multiple tests, buying support and selling resistance offers a systematic approach. The range trader buys near the floor with a stop below it and sells near the ceiling, capturing the oscillation. The strategy fails when a breakout occurs — when the price moves decisively outside the historical range — so position sizing and stop placement are critical. For ML-based trading systems, range detection is often modeled as a regime classification problem: is the asset currently trending or oscillating? Different strategies are optimal in different regimes, and the model's job is to identify which regime applies.
All of these strategies benefit from volume analysis, and the OBV momentum-and-volume indicator is one of the simplest tools for the job. On-balance volume adds today's volume to a running total when the close is up and subtracts it when the close is down. The resulting line tracks whether volume is flowing into or out of a security over time. A breakout from a range that is accompanied by rising OBV carries more conviction than one that occurs on thin volume — high OBV suggests that the move reflects broad participation rather than a low-liquidity anomaly. In pairs trading, OBV divergence between two correlated stocks can sometimes flag the beginning of a spread expansion before it shows up in price alone.
For ML engineers building trading features, these five concepts translate directly into measurable signals: futures curve shape as a feature for commodity models, short-dated option implied volatility changes as an input to regime classifiers, rolling pair-spread z-scores for mean-reversion models, price-channel boundaries for range-breakout detection, and OBV slope as a volume confirmation filter. The underlying strategies have been studied and refined over decades, which means there is a rich literature of backtests and edge cases to draw on when designing training data and evaluation procedures.