Systematic trading research, validation and engineering
How Lucitech moves a systematic trading candidate from structured
discovery through broker-aligned historical testing, independent
evidence gates, controlled deployment and live observation.
Candidate focus: EUR/USD intraday mean-reversion
Signal horizon: 30 minutes
First publication: 19th July 2026
Last revised: 20th July 2026
A persuasive backtest is easy to produce. A defensible decision to
continue researching, approve or cautiously deploy a systematic
strategy is much harder.
This case study documents the evidence lineage of one Lucitech
candidate. It does not disclose the complete strategy, winning
parameter values or a reproducible optimisation recipe. Its purpose is
to demonstrate the process used to decide whether evidence is strong
enough for a candidate to advance—and when it should stop.
A historical result describes what a particular configuration did
against a particular dataset under particular assumptions. It does not,
by itself, establish what will happen next.
If enough configurations are tested, some will look successful by
chance. Selecting the best result without accounting for the wider
trial population can turn luck into an apparently compelling
strategy.
A configuration can become too closely adapted to the precise
historical period used to find it. It may describe past noise
exceptionally well while failing on new data.
A test becomes unreliable if information that would not have been
available at the decision point accidentally influences the signal,
parameter selection or validation process.
Attractive results may disappear when spread, slippage, latency,
financing, fill behaviour and other operating frictions are more
realistically represented.
A candidate may work only during a particular volatility, trend or
liquidity regime. Aggregate performance can conceal long weak
periods or dependence on a small number of favourable episodes.
An extreme profit factor or smooth equity curve can be produced by
very few trades. Small samples provide little evidence about the
range of outcomes the strategy may encounter.
The candidate considered here belongs to a generic intraday
mean-reversion strategy family operating in EUR/USD on a 30-minute
signal horizon.
The first-phase implementation is deliberately bounded. Entries are
evaluated from 30-minute market data and, once a position is opened,
static profit-target and stop-loss levels govern the trade. A possible
later research phase may examine tick-based stop management informed by
maximum favourable excursion and maximum adverse excursion evidence.
That later work is not part of the result documented here.
Lucitech treats optimisation as the beginning of an investigation,
rather than its conclusion. A candidate must move through a recorded
sequence of evidence gates.
Figure 1 — Lucitech’s evidence pipeline records whether a candidate
advances, stops or returns for further investigation at each gate.
Each Optuna candidate is evaluated through Lucitech’s headless JForex
testing path rather than through a detached spreadsheet or simplified
candle-only simulator.
A code-level audit of the runner confirms that the target instrument is
subscribed, historical data is downloaded and the test is configured
with
ITesterClient.DataLoadingMethod.ALL_TICKS
before the strategy is started through the JForex tester client.
This means candidate selection takes place through the project’s
broker-platform testing route using the same core strategy
implementation employed by the deployment path. It provides a more
broker-aligned historical evaluation than a separate simplified
simulator.
Further information:
Dukascopy JForex Historical Tester documentation
.
Optuna helps search a large parameter space consistently. It does not
make the resulting candidate profitable, remove overfitting risk or
replace subsequent validation.
Each trial represents one candidate parameter configuration. The
configuration is evaluated through the headless JForex historical
testing path and its resulting evidence is recorded for screening.
Trial scoring is designed to distinguish potentially useful candidates
from results that are too weak, unstable or statistically thin to
justify further work.
A later maintenance evaluation of the T0094/P95 parameter lineage was
completed on 12 July 2026. Across its six-month historical window it
recorded 546 trades, 170.8 net pips and a profit factor of approximately
1.095. The result provided continuing evidence for an existing
candidate; it was not treated as standalone proof or sufficient
evidence for deployment.
Lucitech publishes enough information to make the research discipline
understandable and reviewable, while withholding the combination of
details that would make the strategy search directly reproducible.
| Area | Published | Withheld |
|---|---|---|
| Framework |
Optuna-led structured parameter discovery and JForex historical evaluation. |
Custom implementation details that would reconstruct the search. |
| Parameters |
Broad parameter categories and the reason those categories are investigated. |
Exact feature definitions, complete ranges and winning values. |
| Scoring |
Broad score components, their purpose and the kinds of fragility penalised. |
Exact formula, weights, thresholds and ranking rules. |
| Results |
Trial populations, distributions, representative results and candidate progression. |
Full trial databases and deployment-ready parameter files. |
| Validation |
Evidence gates, material outcomes and reasons for progression or rejection. |
Details that would reveal or reproduce the executable edge. |
Any individual value may be harmless in isolation. The complete
combination of feature definitions, parameter ranges, objective
function, penalties, thresholds and winning configuration can amount
to a reproducible search recipe. That recipe forms part of Lucitech’s
research and strategy intellectual property.
The historical result establishes a candidate for investigation. The
next question is whether its behaviour remains defensible when the
original search conditions no longer control the test.
The Workbench candidate record links T0094/P95 to a nine-month rolling
maintenance source dated 13 June 2026. Its evidence summary records 406
trades, 159.902 net pips, an average of 0.3938 pips per trade and a
profit factor of 1.0936. It also records a worst-year profit factor of
1.0188, no negative years in that evaluation, and classification as
OOS_PASSED_STRONG on 14 June 2026.
These values remain modest. Their importance is not that they prove a
durable edge, but that the candidate identity, source evidence and
subsequent decision are retained as a traceable record. The later July
maintenance result shown above is a separate evaluation associated
with the same T0094/P95 parameter lineage.
OOS_PASSED_STRONG
On 14 July 2026, the T0094/P95 lineage entered controlled EUR/USD
micro-lot operation alongside its corresponding demo-shadow benchmark.
The deployment uses fixed micro-lot sizing. It does not employ an
undisclosed post-selection rule pack or additional runtime veto layer.
The active controls are part of the base strategy configuration,
including signal conditions, spread control, static stop and target,
and a bounded holding period.
The purpose of this deployment is evidence collection. It allows
Lucitech to compare expected behaviour with shadow and live
observations while strictly limiting financial exposure. Entry into
live micro-lot operation should not be interpreted as a claim of
validated profitability.
Later revisions will extend the record as additional evidence becomes
suitable for publication.
Out-of-sample outcomes, robustness challenges and the reasons the
candidate was permitted to advance—or required to stop.
Comparison of historical expectations, demo-shadow behaviour and
controlled live-micro observations over a meaningful sample.
Runtime availability, heartbeat and market-data freshness,
execution observations, exceptions and evidence drift.
Beyond the Backtest follows one candidate through the evidence
pipeline. A separate research note examines the horizontal dimension:
how the Alpha Workbench records multi-instrument research intake,
preserves batch provenance and separates raw optimisation output from
governed candidates.
Receive occasional Lucitech research and engineering updates as new
validation stages are documented. You may optionally tell us about
your current or proposed systematic trading work. Please do not
include confidential strategy details or proprietary parameters.
Subscription frequency: occasional. Every message should provide an
unsubscribe route. Personal information should be processed in
accordance with the Lucitech privacy notice.