Orderbook Features Without the Orderbook: LOB Alpha for MT5 at 1–15 Minutes

The most useful finding in the 2025–2026 limit-order-book literature is not a new neural architecture — it is a verdict on where effort belongs. A June 2026 study on crypto order books, sampling the top 200 levels at 100ms, concluded that "better input representations matter more than stacking another hidden layer", with a CNN feeding a gradient-boosted ensemble outperforming pure deep learning. For MT5 and MT4 developers, that reframes an entire research agenda. The orderbook-alpha conversation has been dominated by HFT practitioners with co-location and microsecond budgets — the wrong audience for retail execution running at 50–200ms. The right question for everyone else is narrower and more tractable: which LOB-derived features retain enough signal at 1–15 minute horizons to be actionable, and how do you construct them in a market that has no central order book at all?

The Problem: Forex Has No Order Book to Engineer From

Every canonical LOB result — DeepLOB, the queue-imbalance predictors, the multi-level order-flow models — was built on a centralized limit order book: equities on Nasdaq, listed futures, or crypto on a single venue like Binance. Spot forex has no such object. It is a decentralized OTC market where each broker and liquidity provider exposes only its own slice of depth. As one market-depth primer puts it bluntly:

Each broker, liquidity provider, or ECN may only display orders within their own ecosystem, preventing traders from seeing the true overall market depth at any moment.

The consequence is that the headline feature of equity microstructure — a clean, exchange-consolidated book you can read imbalance directly off of — simply does not exist for EUR/USD. What MT5 gives you instead is a set of proxies: tick volume that counts price updates rather than contracts, a broker-specific depth-of-market feed via MarketBook() when the LP provides one, and the bid/ask spread itself as a dynamic state variable. The engineering task is not to read the book. It is to reconstruct the statistics the book would have produced, from the partial signals the platform actually surfaces.

Why the Mid-Frequency Tier Is the Right Battlefield

The reason HFT-grade LOB features are useless to retail EAs is alpha decay, and it is steep. The foundational queue-imbalance result is that book imbalance has strong predictive power over the very next mid-price move and then decays sharply for later changes. That is a signal measured in single-digit events — invisible to anyone executing at 50–200ms. The same pattern recurs in deep models: across 115 Nasdaq stocks, research found that "alpha decay represents a significant challenge, with traditional models such as DeepLOB losing predictive power as the time horizon increases."

The non-obvious move is to stop fighting decay and instead choose a horizon where the residual signal is still harvestable at retail latency. Two 2026 results point the same way. A January 2026 Temporal Kolmogorov–Arnold Network (T-KAN) — which replaces an LSTM's linear weights with learnable B-spline functions specifically to model decay across horizons — reports improved Sharpe at the 5-minute horizon versus the 1-minute horizon relative to a standard LSTM. And the Hawkes-process order-flow work of Anantha et al. (2024) frames the choice explicitly, examining 30-second, 1-minute and 5-minute windows and identifying response horizons that optimize a quasi-Sharpe ratio. The 1–15 minute band is not a compromise; for a non-co-located trader it is where the signal-to-latency ratio is actually favorable.

Constructing the Proxies MT5 Can Actually Produce

Three features carry most of the load, and each has a defensible MT5 construction. The first is microprice, a volume-weighted mid that tilts toward the heavier side of the book:

microprice = mid * (V_ask - V_bid) / (V_ask + V_bid)
// V_ask, V_bid from MarketBook() DOM levels if the LP
// provides depth; otherwise aggregate bid/ask tick
// pressure over a rolling window as a degraded proxy

The second is order-flow imbalance (OFI), the signed change in resting size at the top of book, approximated in OTC FX from tick-spread dynamics. Commercial MQL5 indicators already ship a tick-spread imbalance OFI proxy compatible with forex, indices, gold and CFDs — evidence the proxy approach has practitioner traction, not just academic provenance. The third is tick-volume imbalance, the least trustworthy input but the most available: MT5 tick volume counts price updates, not contracts, making it, in the words of one depth-transparency analysis, a "deeply flawed metric" — yet still the best OFI surrogate the OTC structure permits.

Before committing any of this to MQL5, it is worth validating the imbalance visually against a second data source. Traders can cross-check broker DOM behavior and volume-profile structure using TradingView, whose Depth of Market and Volume Profile tools render the same queue-imbalance and microprice concepts, and whose Pine Script environment lets you prototype a tick-level imbalance calculation visually before porting it to MarketBook() output — treating the chart as where the feature is sanity-checked, not where a pre-decided signal is confirmed.

Multi-Level OFI and the 'Better Inputs' Thesis

Once the proxy is built, the temptation is to feed it into the deepest model available. The recent evidence argues the opposite. Regularized linear models using multi-level OFI features — pooling imbalance across several book levels rather than top-of-book alone — dominate single-level specifications out-of-sample at short horizons, with multi-level approaches improving both in-sample and out-of-sample fit across 1–5 minute windows. The lift comes from the feature construction, not the estimator.

The crypto literature makes the same point from the representation side. The June 2026 study found that a volume representation of the book outperformed a price-only representation across architectures, and that the architecture mattered less than the input. The DeepLOB benchmark itself — 100-event sequences with 40 features (top 10 bid/ask levels × price and volume), now extended into a 2025 Quantitative Finance microstructural guide — established the template, but the 2026 reading of it is that the 40-feature input design carries more of the edge than the CNN-LSTM stack on top. For an EA developer with finite time, that is permission to spend it on imbalance features, level aggregation and clean sampling rather than on chasing a marginally deeper network.

Why 'Real Ticks' Mode Is Non-Negotiable Here

An imbalance feature is, by definition, a function of intrabar order flow — which means the backtest engine's tick path determines whether the feature even exists. MT5's synthetic Every Tick mode reconstructs intrabar price from only four points per minute (OHLC). Any queue-imbalance, OFI or microprice feature computed inside such a bar is not approximated — it is fabricated, because the underlying tick sequence the feature reads was invented by an interpolator. As the MQL5 testing community has documented, "testing the same EA on Every Tick vs. Every tick based on real ticks can produce completely opposite results."

For any sub-5-minute orderbook-feature EA, Every tick based on real ticks is mandatory, sourced from broker-side bid/ask history. This is the prerequisite that most LOB-feature experiments quietly skip, and it is why so many promising proxies fail forward. The constraint also interacts with spread regime. EUR/USD runs 0.1–0.2 pips during the London–New York overlap — a favorable signal-to-noise window for 1-minute imbalance features — but widens to 2–3 pips on FOMC and NFP, where the feature is mostly noise and the strategy should filter. Gold is worse: 1.5–3 point spreads normally, ballooning to 20–30+ points on macro prints, rendering OFI proxies unreliable in event windows. With DXY near 98 in 2026's elevated-volatility regime, wider average spreads raise the cost of acting on any mid-frequency book signal, which tightens the bar a feature must clear to be worth trading.

What Transfers — and What Doesn't — From Crypto and Equities

Because forex lacks a native book, the only way to inherit the LOB research is by transfer, and the 2026 evidence on transferability is cautiously encouraging. A microstructure study spanning 3,417,972 minute-level bars across six cryptocurrencies on Binance spot and perpetuals (August 2025–February 2026), using nine engineered book-and-trade features plus three momentum controls, reported "stable cross-asset patterns" in which book features carried predictive importance. Stable cross-asset importance is the property that makes transfer to liquid FX plausible — if the ranking of which features matter holds across very different assets, it is more likely to hold on EUR/USD than any single asset's coefficients would suggest.

What does not transfer is depth quality. USD/JPY tends to offer higher tick frequency and more structured broker DOM depth than the commodity currencies, making its proxy book materially cleaner; thinly-quoted crosses inherit the same feature formulas with far worse inputs. The lineage here is worth keeping in view — from Glosten–Milgrom (1985) adverse selection and Kyle (1985) strategic trading, through Kercheval–Zhang (2015) and the widely-cited DeepLOB benchmark (2019), the field has pivoted from "can we predict?" to "does it survive execution costs and decay?" The answer is a qualified yes, but the qualification is the whole game: it survives only at carefully chosen horizons, with carefully selected features, on instruments whose proxy book is clean enough to trust.

A Feature-First Workflow for the Mid-Frequency Build

Pulling the evidence together, the build sequence for a 1–15 minute orderbook-feature EA inverts the usual model-first instinct:

1. Pick the horizon before the feature. Decay curves say top-of-book imbalance is spent within a couple of events; target the 1–5 minute band where the Hawkes and T-KAN work shows residual signal survives retail latency.
2. Engineer the input, not the network. Build multi-level OFI, microprice and tick-volume imbalance proxies first. The 2026 consensus is that input representation outweighs model depth — a regularized linear or gradient-boosted model on good features beats a deep net on raw ones.
3. Backtest only on real ticks. Synthetic OHLC mode fabricates the very intrabar flow your feature reads. Real-tick mode, broker-side spread, and a non-zero execution delay are the floor, not an upgrade.
4. Filter by spread regime. Suppress signals when the spread leaves its tight band — overlap-session EUR/USD is tradeable; the FOMC and NFP windows, and gold during macro prints, are where the proxy degrades to noise.
5. Prefer clean-book instruments. Favor pairs with high tick frequency and structured DOM (USD/JPY, EUR/USD) over thin crosses where the same formula ingests worse inputs.

None of this promises a durable edge — alpha decay is a property of the signal, not a bug to be patched, and the OTC book will always be a proxy of a proxy. But the framing dissolves the false barrier that orderbook alpha belongs only to co-located HFT desks. At the mid-frequency tier, with disciplined feature engineering and an honest tick environment, the LOB literature has something an MT5 developer can actually use — provided the work goes into the inputs, where the 2026 research keeps insisting the edge actually lives.