Beyond the Backtest: November 2026 Case Study

Lucitech · Version 0.4 · 5 November 2026

Abstract

Lucitech began with an intraday FX hypothesis and developed a
platform for recording research decisions, deploying identifiable
strategy configurations, supervising portfolio risk and reconciling
execution records. This case study examines the resulting evidence at
the research, deployment and broker boundaries. For one reviewed GBP/USD
four-hour DEMO position, retained records establish the chain from
discovery through configured deployment to broker execution and closure.
A separate EUR/USD LIVE review associated 168 supplied broker positions
with local run records and matched their original fill and final closing
prices. That population recorded a net loss of USD 20.74 after reported
commission. Coverage and accounting exceptions remain documented. The
findings support scoped operational traceability and motivate further
alpha and portfolio research.

Keywords · systematic FX · research provenance ·
portfolio supervision · execution reconciliation

1. Introduction

The project began by asking whether intraday price displacements
could be traded systematically. Operating the early implementation
exposed another question. Could an operator identify the research behind
a running configuration, explain how portfolio controls treated an order
and reconcile the result against broker evidence?

The resulting Workbench gives operators access to evidence produced
by a command-line research pipeline, a broker-platform execution layer
and a persistent operations store. Its development reflects experience
in electronics, real-time systems and institutional trading engineering.
The current priority is to develop trading alpha and the portfolio
controls needed to evaluate it. Selected software components and
integration work may be offered where there is a fit with a client’s
existing research, data or execution environment.

2. Background and
conceptual framework

2.1 Search, selection and
qualification

Optuna supplies a programmable optimisation framework [1]. Its use
enables systematic search but does not establish the validity of
selected configurations. Backtest-overfitting research identifies the
risks of selecting attractive results from repeated historical trials
[2]. This motivates retaining search-population context and separating
discovery-exposed observations from previously unseen evaluation. This
study does not currently report a calculated probability of backtest
overfitting or a multiple-testing-adjusted significance result.

2.2 Signal processing
and economic hypotheses

For positive observed prices P_t, define log returns as

r_t=\log(P_t)-\log(P_{t-1}). \tag{1}

(1)

A conceptual representation is r_t=s_t+epsilon_t, where s_t is a
hypothesised predictable component and epsilon_t a residual. This
decomposition is not uniquely identifiable from observations without
additional assumptions.

An exponential moving average provides a direct connection to causal
filtering:

m_t=\alpha P_t+(1-\alpha)m_{t-1},\quad 0<\alpha\leq1. \tag{2}

(2)

A displacement feature may be represented by z_t=(P_t-m_t)/v_t, with
v_t>0 an explicitly defined price-scale volatility estimate. The
exact production calculation, bar timing, warm-up and volatility
estimator must be tied to the implementation before publication. Mean
reversion hypothesises that some eligible displacements subsequently
contract sufficiently to exceed trading costs. The existence of a
filtered displacement alone does not imply predictability.

A possible future diagnostic is the discrete Fourier transform of
returns,

R_k=\sum_{t=0}^{N-1}r_t\exp(-\mathrm{i}2\pi kt/N), \tag{3}

(3)

provides a potential future diagnostic of frequency content. Fourier
analysis is a proposed experiment. The documented pipeline uses
time-domain filters such as the exponential moving average, and the
reported experiments did not use spectral features to select, size or
execute trades. Spectral peaks would require investigation for
stability, window sensitivity and economic usefulness. Full-window
retrospective transformations must not introduce future observations
into trading decisions.

2.3 Non-stationarity and
future learning

Systematic decisions follow explicit rules; uncertainty remains in
the outcomes. Financial observations need not share an invariant
distribution across time. Accordingly, historical fit, temporal survival
and forward observation answer distinct questions.

Bayesian updating provides a possible future inferential
framework:

p(\theta\mid D)=\frac{p(D\mid\theta)p(\theta)}{p(D)}. \tag{4}

(4)

Stage 0 uses Optuna’s Tree-structured Parzen Estimator (TPE) in
documented campaigns. TPE uses earlier trial scores to propose parameter
values and is a form of Bayesian optimisation. Its search model concerns
the backtest objective. Estimating uncertainty about future strategy
profitability would require a separate statistical model and evaluation
protocol. Future adaptive models would require explicit assumptions
about changing parameters or regimes; accumulating data alone does not
resolve non-stationarity.

3. System and evaluation
design

Headless Java/JForex jobs execute historical experiments and the
strategy runtime. Python tools manage search, screening and analysis.
PostgreSQL retains results and telemetry. The Workbench presents
selected evidence and operator controls. The public demonstrator uses
sanitised snapshots. The private operational Workbench is the
environment for inspecting candidate records, linked curves, portfolio
membership and exceptions.

Strategy parameters and deployment policy are distinct contracts. The
strategy contract identifies the build, instrument, timeframe, direction
and parameter vector. The deployment contract identifies the runtime,
account environment, sizing, risk policy and approval. Runtime telemetry
records decisions, orders, fills and exceptions using identities that
can be mapped to broker records. A new strategy or execution integration
requires an explicit mapping and acceptance checks.

The evaluation examines four questions. Can research identity survive
promotion? Can a portfolio decision be explained? Can a specific broker
execution be linked to its local records? What trading outcomes were
observed under a stated measurement scope? Historical replay, DEMO and
LIVE populations are reported separately. The available broker export
dates define the observed LIVE population. Net profit factor above 1
remains one forward evaluation objective. A separate portfolio research
stream is developing a wider scorecard covering net P&L after costs,
trade count, drawdown, exposure overlap and allocation outcomes, with
downside-risk measures such as Sortino to follow once a suitable return
series is available. Alpha discovery and portfolio risk are the main
development priorities. For prospective comparisons, predefine the
observation window, cohort, sizing policy and evaluation criteria.

4. Formal progression model

For a fixed campaign, define the advancing candidate population
by

\boxed{\begin{aligned}\mathcal C_n &=(\mathcal F_n\circ\cdots\circ\mathcal F_1) \bigl(\mathcal D_0(H,\Omega,D_0)\bigr), \\ \mathcal F_k(X) &=\{(c,E\oplus e_k):(c,E)\in X, \\ &\quad e_k=\mathcal T_k(c,E;D_k),\ G_k(c,E\oplus e_k)=\mathrm{PASS}\}.\end{aligned}} \tag{6}

(6)

Here H denotes the hypothesis, Omega the permitted search space, D_k
the identified stage data, D_0 calligraphic the discovery operation, T_k
the experiment, G_k its versioned predicate, E accumulated evidence and
the append operator preserves prior history. The functional stage index
is not asserted to match every existing script label.

Equation (6) specifies admission to subsequent research stages. It
does not specify deletion: all valid failures, pending evaluations and
execution errors remain retained. Outcomes are typed as PASS, FAIL,
PENDING, ERROR or NOT_EVALUATED. Research eligibility is separate from
permission to deploy or allocate capital.

For an elementary economic screen,

G_k=\mathrm{PASS}\iff V_k\land N_k\geq N_{\min,k}\land\Pi_k>0 \land\operatorname{finite}(\mathrm{PF}_k)\land\mathrm{PF}_k>1, \tag{7}

(7)

where V_k expresses evidence validity, N_k closed-trade count and
Pi_k net trading P&L. Stage-specific requirements may be stricter.
Passing is a deterministic statement about recorded observations and
rules, not certainty about future outcomes.

Operational preconditions include identified inputs, preceding
decisions and a frozen configuration. Postconditions include a retained
result, provenance, gate outcome and reasons. Intended invariants are
identity preservation, traceability, history preservation, controlled
material changes and explicit authorisation. These have not been
formally verified against the implementation; the notation is a
specification, not a proof of software correctness.

5. Findings

C07 identifies a GBP/USD 4H cluster formed during post-processing of
the discovery results. In this paper, C07 also refers to the Stage-1
representative selected from that cluster and carried forward as the
frozen candidate.

Evidence	Reviewed result	Scope
GBP/USD 4H C07	Research configuration traced through deployment, execution and closure	One reviewed DEMO position
DEMO execution checks	85 original fill prices and 82 comparable exit prices matched	Linked subset of supplied account history
EUR/USD LIVE execution checks	168 one-to-one position associations and matching original fill and final closing prices	Supplied 15 July–4 November 2026 history
EUR/USD LIVE trading outcome	USD −20.74 after reported commission, net cash PF 0.7508	Same 168 supplied positions
AUD/USD 4H historical research	446 closes, USD +101.53 after settled fees, PF 1.1522	Selected unchanged historical replay, 7 October 2021–7 October 2026

For the reviewed GBPUSD C07 position, retained evidence traces the
research configuration through configured deployment to broker
execution, including exact position/label and member/run/reservation
linkage, quantity agreement, original fill-price agreement and
subsequent closure.

Early research used fixed order amounts to compare strategy
configurations under a common sizing basis. The newer portfolio layer
adds supervised admission, order allocation and sizing informed by stop
distance and the applicable risk budget. Position-level stop rules and
portfolio limits serve different roles. The platform also has
configurable guardrails, including a maximum daily loss. Their
activation and enforcement must be reported for each reviewed
deployment. The allocation demonstrator exposes full, reduced and
blocked decisions. Reviewed records support explanations based on entry,
currency-pair, portfolio and directional currency constraints. These
records explain the allocator’s recorded decision. Execution is
evidenced separately by broker records. Observe-only records describe
what the control would have done.

Attributed LIVE run	Positions	Gross USD	Commission USD	Net USD	Net cash PF
EUR/USD 30M P95	153	−3.68	−12.24	−15.92	0.7755
EUR/USD 1H	15	−3.62	−1.20	−4.82	0.6084
Total	168	−7.30	−13.44	−20.74	0.7508

The LIVE account classification comes from the operator and runtime
evidence. The supplied CSVs do not contain an account number. The table
describes closed position outcomes, including reported lifecycle
adjustments. An equity curve and account drawdown require a reconciled
valuation series and cash-flow baseline.

The EUR/USD 30-minute P95 LIVE sample recorded 153 closed positions
and net P&L of USD −15.92 after reported commission. A separate DEMO
shadow snapshot showed 136 trades and a displayed profit factor of
1.1088. The LIVE spread cap can block entries that the DEMO shadow
takes. Position-ownership blocks have also been recorded. Quantifying
each effect requires matching signals, admission decisions and
executions across the two environments.

6. Discussion and limitations

Early operation drove the addition of explicit lineage, ownership,
allocation and reconciliation evidence. C07 shows how those records can
be followed for a newer portfolio member. P95 remains useful as a legacy
comparison. Its discovery and recorded evaluation trade spans overlap,
and the declared historical partitions were unavailable in the reviewed
records. Its later histories should retain their own identifiers and
qualifications.

The selected AUD/USD candidate finished the five-year historical
replay positive overall but lost money in the earliest annual slice. The
Workbench’s stage selector lets readers inspect the curves by period.
Overlapping research windows and limited samples constrain what these
historical continuation checks establish. Forward DEMO observation is
tracked separately.

An operator can inspect evidence, identify a control decision and
investigate a discrepancy. In a client deployment those capabilities
could feed internal audit, governance and reporting workflows.
Client-specific adapters, field mappings and operating responsibilities
require agreement and testing. The study evaluates evidence and
integration capability. Regulatory report submission and compliance are
outside its evaluated scope.

7. Conclusion and next work

The platform has progressed from isolated experiments to a documented
research and operating workflow with scoped broker reconciliation. The
current priority is alpha research and portfolio risk development.
Planned work includes prospective DEMO observation, exposure and
drawdown overlap analysis, comparison of sizing policies and separately
identified trend-following experiments. A development goal is to run
multiple portfolios concurrently, each with its own membership,
allocation policy and risk limits. The supported portfolio count will be
measured under representative load. Compute and storage capacity,
database throughput and latency, and the number of active strategies and
volume of telemetry per portfolio will shape that limit. Shared
broker-account constraints also apply where portfolios use the same
account.

Spectral diagnostics, Bayesian learning and managed research agents
remain research directions. Public evidence can document progress while
the private Workbench supports daily investigation.

Appendix A. Worked
evidence and exceptions

A.1 C07 lineage

Discovery trial → unchanged counterpart replay → Stage-1 cluster
representative → frozen candidate → configured deployment → reservation
and broker execution → closure.

The audit matched all 14 research fields across the retained records
and configured deployment. It compared source and parameter
fingerprints. Supervisor sizing was identified separately from the fixed
research baseline. The broker position ID and the entry and exit
execution-order IDs occupy distinct namespaces. The exact associations
are retained privately. The reviewed position establishes identity,
quantity, original fill price and closure. Independent stop-risk
arithmetic and complete position-level cost attribution remain open.

A.2 DEMO reconciliation

The two detailed exports contain 1,831 distinct execution and
rollover rows across seven pairs, representing 849 identified broker
positions. The local review found 85 PostgreSQL-linked positions and
five explicit ledger-only probes. Fifty-four of the 85 had
reservation-backed member/run and quantity evidence. The other 759
broker positions lacked a local identity match in the reviewed sources.
A reverse census produced 435 local candidates without confirmed account
attribution. The probes remain in account cash and are separated from
strategy performance.

For the 22 October–1 November review interval, under a tested
reporting-day convention, aggregate gross USD +25.69 and commission USD
−6.63 bridge to USD +19.06 of reported account movement. This is
whole-account cash, including activity outside the four-hour portfolio.
Two opposing one-cent date-group differences remain visible. Wider
completed-date conversion differences total USD −0.11 and provisional
valuation differences total USD −1.58. Broker cutoff and unsettled mark
conventions remain open.

A.3 LIVE reconciliation

The detailed report has 366 rows. The summary has 354. Ten
split-close rows and two position-attributed rollover rows explain the
difference. The 336 ordinary execution orders agree after fill
aggregation. An initial numeric lookup missed textual legacy order keys.
Exact position and execution IDs co-recorded in receipt messages
resolved the associations, including ten initially ambiguous
timing/price candidates.

One local cash value was missing and six rollover positions had
one-cent local gross-cash differences. The two broker views differed by
one cent overall. Local filled-quantity fields were unavailable, while
broker quantities balanced within the supplied lifecycles. Independent
account cash statements and an open-position snapshot are outstanding.
Original-fill events and later adjusted entry fields should remain
separately labelled.

A.4 Supplementary AUD/USD
research

Discovery window	Completed trials	Own-window passes	Also pass extra one pip per close
12 months	600	256	164
36 months	600	309	197

The windows overlap. C06 failed its matched 12-month screen and was
parked. C12 passed five of six separately initialised half-year slices
and the frozen five-year historical screen. The five-year replay
recorded USD +56.93 and PF 1.0834 after a post-processing one-pip charge
per close. The first anniversary interval lost USD 32.60. The expanded
history remains retrospective.

Appendix B. Calculation
conventions

B.1 Net outcomes and cost
sensitivity

For trade i, let x_i be the price P&L less allocated fees and
financing included under the declared accounting policy. Spread and
execution effects already embedded in fill prices must not be subtracted
twice. Define

\Pi=\sum_i x_i,\qquad \mathrm{PF}_{net}=\frac{\sum_i\max(x_i,0)}{\sum_i\max(-x_i,0)}. \tag{8}

(8)

A zero denominator is reported as undefined/no observed losses, with
trade count; it is not treated as compelling evidence. Cash PF and pip
PF must be distinguished where sizes vary. Net trading results exclude
business infrastructure overhead unless explicitly added.

For an extra cost delta pips per closed trade and pip cash value
w_i,

x_i^{(\delta)}=x_i-\delta w_i. \tag{9}

(9)

For the fixed-size AUD/USD experiment, w_i=USD 0.10/pip. This
arithmetic stress changes outcomes on unchanged trades. It does not
model altered fills, stops, liquidity, order rejection or signal
paths.

B.2 Drawdown and uncertainty

For a positive marked-to-market equity series E_t and running maximum
M_t,

\mathrm{MDD}_{\%}=\max_t\frac{M_t-E_t}{M_t},\quad M_t=\max_{u\leq t}E_u. \tag{10}

(10)

Cash drawdown on cumulative closed-trade net is a different statistic
and is labelled accordingly. Deposit-normalised returns can be
misleading where the research deposit is arbitrary relative to fixed
micro-lot risk.

No confidence intervals are supplied in this draft. Before
performance conclusions, inspect aligned observations and choose an
uncertainty procedure respecting serial dependence and shared signals. A
block-resampling sensitivity analysis is a candidate, not an analysis
already performed. It cannot by itself correct the entire historical
selection process.

Research PF reported by JForex is described by the operator as net of
its accounted costs. For the legacy P95 specimen, the exported pips
reproduce the displayed research PF to four decimals. The exact SDK
field, version and cost mapping require documentation. Cash PF is
reported separately. The public chart’s daily closed-P&L drawdown
uses cumulative daily pips from a zero baseline and excludes
open-position valuation.

References and evidence

[1] Akiba et al. (2019). Optuna: A Next-generation Hyperparameter
Optimization Framework. https://doi.org/10.1145/3292500.3330701

[2] Bailey et al. The Probability of Backtest Overfitting.
https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf

Internal evidence includes the AUD/USD discovery and five-year
reviews (30 October), the C07 lineage audit (1 November) and the DEMO
and LIVE broker reconciliation drafts (4 November). Publication extracts
should be sanitised and pinned to dated evidence versions. The author
has a commercial interest in Lucitech. This manuscript was prepared with
AI assistance and awaits author verification.

Beyond the Backtest