Adaptive Intelligence: How PMTS Re-Trains Its Machine Learning Models on New Market Data

Markets are non-stationary. The statistical relationships that defined gold's behaviour during one volatility regime can decay or invert in the next, and a model trained once and frozen will quietly degrade as the distribution it learned drifts away from the live tape. For an institutional trading system, the question is not whether to use machine learning, but how to keep the models honest as conditions change. At PMTS, model re-training is not an occasional maintenance task — it is a governed, continuous process with explicit validation gates that decide whether a newly trained model is ever allowed to touch live capital. This article walks through how that pipeline works, and the live KPIs that demonstrate it is doing its job.

As of June 18, 2026, the reference strategy tracked on our public dashboard shows a win rate of 86.79% across 53 trades, a profit factor of 6.53, and a Sharpe ratio of 10.42, with maximum drawdown contained to 0.41% and a cumulative return of 10.81%. These figures are not the product of a single lucky calibration; they are the output of a retraining discipline designed to adapt without overfitting.

Why static models fail in production

A trading model encodes a hypothesis about the joint distribution of features — volatility, order-flow imbalance, term-structure of real yields, session seasonality — and the forward return of an instrument such as XAUUSD. That distribution is conditional on a macro regime: the reaction function of the Fed, the path of real yields, positioning around scheduled events like FOMC. When the regime shifts, the conditional distribution shifts with it. Three failure modes follow.

• Covariate shift — the input features move into ranges the model rarely saw in training, so its predictions become extrapolations rather than interpolations.
• Concept drift — the relationship between features and forward returns itself changes, so even well-sampled inputs map to the wrong output.
• Label decay — the economic meaning of a signal erodes as the behaviour that generated past edge gets crowded out or arbitraged away.

A frozen model has no defence against any of these. The answer is not to retrain blindly on every new tick — that simply chases noise — but to retrain on a disciplined cadence under a validation regime strict enough to reject models that have learned the wrong thing.

The PMTS retraining pipeline

1. Rolling data windows

Training data is assembled on a rolling window rather than an ever-growing archive. Older observations are down-weighted or dropped so the model preferentially learns from the regime that is actually live, while a longer reference window is retained to preserve memory of rare but recurrent stress events. The result is a model that is current without being amnesiac. All features are reconstructed from the same tick-level MetaTrader 5 feed that drives live execution, so there is no gap between the data the model is trained on and the data it sees in production.

2. Walk-forward validation

Every candidate model is evaluated with walk-forward analysis: it is trained on an in-sample window, then tested on the immediately following out-of-sample period it has never seen, and the window is advanced repeatedly across history. This is the single most important guard against the curve-fitting that makes a backtest look spectacular and a live account bleed. A model that only performs on the data it was optimised against is rejected at this stage, regardless of how attractive its in-sample metrics appear.

3. Drift detection and triggering

Retraining is triggered by two complementary clocks. The first is a scheduled cadence that guarantees no model goes stale past a fixed horizon. The second is event-driven: statistical monitors watch the live feature distribution and the model's realised error, and when either crosses a threshold — a divergence between predicted and realised outcomes, or a feature distribution that has visibly moved — an out-of-cycle retraining is requested. This means the system reacts to regime change rather than waiting for the calendar.

4. Promotion gates

A newly trained model does not replace the incumbent automatically. It must clear quantitative promotion gates measured on out-of-sample data: minimum out-of-sample Sharpe, a profit factor above a defined floor, a maximum-drawdown ceiling, and consistency of performance across sub-periods rather than reliance on a handful of outsized trades. Frameworks such as Sortino and Calmar sit alongside Sharpe in this gate so that downside volatility and drawdown-adjusted return are evaluated explicitly, not just raw volatility-adjusted return. Only a model that beats the incumbent on these criteria is promoted; otherwise the existing model continues to trade.

From model to live execution

Once promoted, a model is deployed into the same execution environment that produces the numbers on our dashboard. Signals are generated, risk-sized, and routed to MT5 under fixed position-sizing and stop discipline, so the edge the model was validated on is the edge that is actually expressed in the account. The transparency of that chain matters: every trade the system takes is recorded and surfaced, which is what allows the retraining process to be held accountable against live results rather than marketing claims. You can review the live performance on the PMTS performance dashboard.

What the live KPIs tell us

The point of this entire architecture is not elegance for its own sake — it is realised, repeatable performance. The reference account's profit factor of 6.53 means gross profit exceeded gross loss by more than six to one over the sample; the Sharpe ratio of 10.42 reflects returns delivered with tightly controlled variance; and a maximum drawdown of 0.41% indicates that adaptation has not come at the cost of risk discipline. At the portfolio level, aggregated managed strategies recorded monthly returns ranging from +7.96% to +30.03% in June 2026, with one core strategy posting +10.90% on the month — a spread that reflects different risk mandates rather than inconsistency in the underlying engine.

It is worth stating plainly what these numbers do and do not mean. A high win rate and a low drawdown over a defined sample are evidence that the retraining discipline is currently well-calibrated to the prevailing regime. They are not a promise about the future, and no serious quantitative process treats them as one. The retraining pipeline exists precisely because the future will differ from the past, and the system's job is to keep adapting as it does.

The institutional case for continuous retraining

For a capital allocator, the relevant due-diligence question is not "what return did the model produce?" but "what process produced that return, and will it survive a regime change?" A governed retraining pipeline — rolling windows, walk-forward validation, drift-triggered updates, and quantitative promotion gates — is an answer to the second question. It converts machine learning from a one-off optimisation into an ongoing discipline, and it is the reason the live KPIs above have remained stable rather than decaying. If you want to evaluate the system against your own criteria, you can open a PMTS account and follow the strategy in real time.

Past performance does not guarantee future results. Trading involves substantial risk of loss and is not suitable for every investor. The KPIs cited reflect a specific reference period and managed accounts with differing risk mandates; individual results will vary.