Quantitative finance · 2026

Chronos

Research, test and inspect a trading idea in one place.

The trading work has grown through several application builds, from early stock analysis and an Electron terminal to a Windows desktop platform with local forecasting services. The aim is to make a strategy easier to inspect: what it saw, why it acted and how the result compares with the original research.

Status
Research platform · Training phase
Domain
Trading research · Desktop applications
Role
Application development, strategy research and model integration
Year
2026

Chronos · AI model concept

From input to
an output worth exploring.

Follow a signal through an encoder, connected hidden layers and an output head. The network and surfaces are illustrative, with no live training data or measured results.

InputsEncoderHidden layersOutput headSynthetic surface

Input encoding

Sample inputs enter a shared representation. The first layers light up as signals move through the network.

Static illustration

Chronos · The defined training programme

Training a model around the whole decision.

The architecture is defined, and Chronos is in the training phase. A compact market forecaster sits within a broader evaluation system: time-aware data, calibrated uncertainty, realistic costs and a record of every experiment.

Parameter design range
1–10M
At each forecast step
5 quantiles
Forecast horizon
50 bars
The full technical foundation
27 chapters

The model inside the decision process

  1. 01

    Prepare

    Closed bars and information available at the decision time.

  2. 02

    Represent

    Normalised windows, patches and explicitly typed covariates.

  3. 03

    Forecast

    A shared transformer with instrument-aware quantile outputs.

  4. 04

    Calibrate

    Adjust uncertainty and inspect the current market regime.

  5. 05

    Decide

    Apply costs, portfolio constraints and a recorded reason to act or wait.

01 · Data and architecture

Give the model a clear view of what it could know.

Market data, timestamps and feature contracts

The specification covers 21 instruments across seven timeframes. Each training example carries its context window, forecast horizon and the availability of every external input. Price history, publication dates and revised economic series need different treatment.

  • Static inputs describe the instrument or timeframe. Known-future inputs describe calendars and scheduled events. Observed-past inputs enter only after their release timestamp.
  • Point-in-time joins and vintage-aware economic data prevent a later revision or delayed release from appearing in an earlier decision.
  • Window normalisation uses context data only. Overlapping forecast labels are accounted for in sample weighting and in train, validation and test splits.
The custom quantile transformer

The defined architecture combines patch-based time-series modelling with a shared backbone and a dedicated layer for relationships between instruments. A compact reference configuration uses four transformer blocks at width 256, within a one-to-ten-million-parameter design range.

  • Reversible instance normalisation puts each window into its own scale; temporal patches turn sequences of bars into manageable tokens.
  • Channel-shared processing learns across instruments. Variate attention provides an explicit, testable place for cross-instrument information to enter.
  • Typed covariates pass through gated variable selection. The output head produces the 10th, 20th, 50th, 80th and 90th percentiles at each of 50 future steps, then returns them to price units.
Baselines, foundation models and adaptation

A custom architecture needs a useful comparison. The programme includes random-walk and linear baselines, volatility models, boosted-tree probes and foundation-model forecasts before attributing an improvement to the transformer.

  • PatchTST, iTransformer and the Temporal Fusion Transformer inform different parts of the design: temporal patches, variate attention and typed covariates.
  • TimesFM and the separate Chronos foundation-model family are reference architectures. The Chronos platform described here is the Luna Labs application and research programme.
  • Low-rank adaptation remains a research comparison. The selected training path is the smaller purpose-built model, with shared representations and market-specific heads.

02 · Training and uncertainty

Train the distribution, then check what its uncertainty means.

Objectives, optimisation and checkpoint selection

The training recipe is defined around quantile forecasting rather than a single predicted price. Pinball loss trains the quantiles; CRPS assesses the predictive distribution, and the programme also specifies a decision-focused objective that incorporates trading costs.

  • AdamW, a warmup-and-cosine learning-rate schedule, global-norm gradient clipping and recoverable checkpoints form the optimisation recipe.
  • Early stopping follows held-out distributional quality. Learning rates, context lengths and other run choices belong in the experiment record.
  • Multi-task sharing is evaluated for negative transfer. A shared model does not justify sharing a market’s calibration or cost assumptions with another market.
Calibration, volatility and regime context

Forecast intervals need checking as market conditions change. The specification combines conditional-width models with conformal calibration and an adaptive coverage controller, then evaluates the result within meaningful market regimes.

  • HAR and GARCH provide interpretable width baselines. Adaptive conformal methods update the treatment of forecast errors as completed outcomes become available.
  • Coverage, interval width, quantile ordering and probability-integral-transform diagnostics are reviewed together; one aggregate coverage number is insufficient.
  • Hidden Markov models and change-point methods supply regime context. Decision-time state estimates must use past and present observations, with hindsight-smoothed states reserved for offline analysis.
  • Conformal guarantees depend on their assumptions. The review explicitly examines dependence, changing regimes and coverage within stressed periods.

03 · Evaluation and decisions

Make every added component earn its place.

Validation, ablations and the experiment ledger

The evaluation design treats the search process as part of the experiment. It records unsuccessful variants as well as selected candidates, and tests whether an added feature improves results after the same costs and constraints are applied.

  • Purging and embargoes separate overlapping labels. Combinatorial purged cross-validation produces several out-of-sample paths; a reserved sequential walk-forward provides a later confirmation step.
  • Ablations isolate the contribution of cross-instrument attention, covariates, regime context and added model complexity against simpler alternatives.
  • The Deflated Sharpe Ratio and probability of backtest overfitting address selection across repeated trials. Forecast comparisons and calibration checks answer separate questions about prediction quality.
  • Deliberate leak-injection checks exercise the evaluation harness. The ledger retains data windows, configuration, model version, costs and the full trial count.
Execution costs and the option to do nothing

The decision layer works with expected outcomes after costs. A measured cost surface distinguishes instruments, session hours, order size and market conditions, so a model is not rewarded for opportunities that disappear at the spread.

  • Spread, slippage, financing and rollover belong in the setting where they occur. Mid-price candles alone cannot establish execution costs.
  • The research sequence compares envelope-based policies with a volatility-targeting baseline and includes a higher-cost stress case.
  • A pre-agreed minimum-edge condition can return no order. The recorded output includes the reason for rejection as well as the reason for a proposed trade.
Position limits, portfolio exposure and tail risk

The specification separates a forecast from permission to take risk. A sizing stack combines a discounted Kelly ceiling, volatility targeting, a regime adjustment and a conditional-value-at-risk cap, alongside portfolio-level exposure checks.

  • The tightest applicable constraint limits the proposed position, and the binding constraint is recorded for review.
  • Covariance, concentration and changing hedge relationships matter across positions. Several individually sized trades can still express the same underlying exposure.
  • Tail assumptions are named: five stored quantiles do not describe every extreme outcome. Gap risk, margin, option convexity and contract mechanics need their own treatment.

04 · Runtime and expansion

Keep training, review and execution connected.

The model artifact and the decision loop

The deployment specification brings the trained model back into the C# platform. Python supports research and training; a frozen ONNX model is the target for native inference, while versioned forecasts, calibration state and cost data support inspection.

  • The existing integration uses Python forecasting services and SQLite outputs read by the desktop application. The custom-model path defines how that boundary can become a frozen inference artifact.
  • Each closed bar follows a defined sequence: read eligible data, forecast, calibrate, assess the regime, apply risk and cost constraints, then record the decision.
  • Training and model selection stay outside the execution loop. A retrained candidate returns through evaluation before replacing a reviewed artifact.
Scenario generation and additional markets

The wider design includes diffusion-based path scenarios and signature-based generative methods for stress testing. These sit in an offline or asynchronous research lane, separate from the direct signal path.

  • Generated paths support tail analysis and scenario comparison. Their value is tested against simpler risk models, with an explicit compute and latency budget.
  • The market roadmap moves from the primary CFD research into options, equity pairs and commodities. Each has distinct contracts, financing, liquidity and risk mechanics.
  • Representations can be shared, while market heads, calibration data, cost models and confirmation runs remain specific to their setting. Joint calendar splits guard against leakage across markets.

05 · Research sequence

A defined route from training to a reviewed artifact.

These are the programme’s review stages. They describe the criteria and dependencies around the training work, rather than a list of completed approvals.

  1. 01

    Prepare the evidence

    Measured costs, point-in-time inputs and checks that can detect leakage.

  2. 02

    Establish the baseline

    Compare existing forecast envelopes with simpler sizing policies after costs.

  3. 03

    Improve the uncertainty

    Test conditional volatility and adaptive calibration by market regime.

  4. 04

    Test useful context

    Evaluate regime routing and covariates before assigning their value to a larger model.

  5. 05

    Train and compare

    Train the bespoke transformer; compare it with the strongest simpler candidate.

  6. 06

    Review for promotion

    Complete the reserved confirmation run and artifact review before deployment changes.

What the review keeps in view

The programme tracks cost dominance, data leakage, repeated-trial selection, calibration drift, correlated positions and unmodelled tails. Each stage has a decision written in advance, including retaining a simpler model when the added component does not help.

06 · Technical coverage

The architecture, the mathematics and the market.

The programme spans model architecture, its mathematical foundations and the market conditions around each decision. The topics below show the breadth of the training and evaluation design.

Model design · Statistical evaluation · Market mechanics

Part I · The machineArchitectures, training and model families
  1. 01What a Neural Network Is
  2. 02Training: How Networks Learn
  3. 03From Sequences to Tokens
  4. 04Attention from First Principles
  5. 05The Transformer, Assembled
  6. 06Transformers for Time Series
  7. 07Foundation Models: TimesFM and Chronos
  8. 08Adaptation: Fine-Tuning, LoRA, and Many Markets
  9. 09Generative Models and Auxiliary Architectures
Part II · The mathematicsThe tools behind the training and evaluation
  1. 10Linear Algebra Beyond the Baseline
  2. 11Multivariable Calculus and Optimisation
  3. 12Probability from First Principles
  4. 13Statistics: Estimation and Inference
  5. 14Quantiles, Scoring Rules, and Calibration
  6. 15Conformal Prediction
  7. 16The Mathematics of Volatility
  8. 17Markov Chains, Hidden States, and Change-Points
  9. 18Information Theory and Learning Theory
  10. 19Backtest Statistics I: Sharpe, PSR, DSR, PBO
  11. 20Backtest Statistics II: Validation, Forecast Tests, Leakage
Part III · The marketTrading mechanics and the assembled platform
  1. 21Prices, Returns, and the Stylised Facts
  2. 22Market Plumbing: CFDs, Spreads, and the Cost Surface
  3. 23Position Sizing: Kelly, Vol Targets, and CVaR
  4. 24Portfolio Mathematics, Pairs, and the Kalman Filter
  5. 25Options from First Principles
  6. 26Commodities, Carry, and the Versioned System
  7. 27Capstone: The Platform, Reassembled
Supporting reference materialSymbols, terminology, formulas, references and platform summary
  • Master Symbol Glossary
  • Term Glossary
  • Formula Quick Reference
  • Annotated References
  • The Platform Canon

Stack

  • C#
  • .NET 8
  • WPF
  • WebView2
  • Python
  • Rust · GPUI
  • C++
  • SQLite
  • TimesFM

An evolving trading research platform connecting strategy experiments, backtesting, forecasting and a desktop view of paper-trading decisions.

01

An evolving platform

The project folders record several generations: stock analysis tools, a Python-backed Electron terminal, a Rust and GPUI shell, a strategy engine with a Tauri interface, and a later C# Windows terminal.

These are related builds with different roles, bringing research, execution and visual inspection together over time.

02

Follow the decision

The current terminal includes a flow view linking strategies to signals, position sizing, open trades and exit monitoring.

Charts and diagnostic views make rejected signals and model forecasts available for inspection alongside the trading state.

03

Keep forecasting separate from the interface

Local forecasting services provide TimesFM and other model outputs to the application.

Separating those processes lets the interface remain responsive while forecast work runs and gives each service its own operational checks.

04

Research that records its limits

Backtests and paper-trading runs are experimental evidence.

The accompanying research programme covers costs, repeated trials, calibration and pre-agreed stopping rules, so an attractive result can be tested before it is treated as a useful strategy.

Pipeline

Trading research and paper-trade review

Stage Can reject

  1. 01

    Market data

    Research inputs

  2. 02

    Forecast

    Local model outputs

  3. 03

    Strategy

    Signals to inspect

  4. 04

    Risk checks

    Sizing and limits

  5. 05

    Paper trade

    Experimental decisions

  6. 06

    Review

    Charts and decision history

    Current Windows terminal uses WPF

The mathematics

A closer look
at the method.

The ideas behind the analysis, with the main terms explained. These are method summaries; fitted parameters and project-specific calculations remain private.

01

A position-sizing research example

One strand of the strategy research combines weighted conditions with a volatility-based size calculation and a drawdown throttle. This illustrates that research idea, rather than specifying every version of the trading platform.

κ = Σg wg · 1[g] Q = κ·R·E λ·ATR · θdd
κ
Conviction score from the weighted gates
1[g]
Indicator that gate g passed
R
Risk fraction per trade
E
Account equity
ATR
Average true range — the volatility unit
λ
Stop distance in ATR units
θdd
Drawdown throttle, 0 when a breaker has tripped

The volatility term scales the intended exposure, and a zero drawdown multiplier suppresses the proposed size. Realised risk also depends on instrument units, gaps, execution and costs. This schematic omits implementation details and is not evidence of trading performance.

Research results

Earlier strategy research: two scenarios, one sample

This earlier two-week foreign-exchange backtest compares two selections from the strategy engine. It predates the custom-transformer training programme above. S2 prioritised raw return; S3 used a lower-correlation selection of strategy families. The return column annualises this short sample. Net AUD shows the result for two account sizes, B and A. These are backtest results, not a live track record.

# Scenario Trades Net AUD (B) Net AUD (A) Return (ann.) Win rate Max DD PF Sharpe
S2 Max return 131 6,581.73 87.69 205% 61.8% 1.61% 2.32 12.67
S3 Independent families 120 6,563.77 86.79 202% 59.2% 0.30% 3.45 17.92

In this sample, S3 kept a similar return while recording a lower drawdown: 0.30% compared with 1.61% for S2. That makes the strategy-selection method worth investigating further. Two weeks is too short to establish whether the difference will persist, and annualising the return does not make it a one-year result.

Behind the work

The skills
that came together.

Research

  • Strategy and backtest development
  • Forecast integration
  • Risk and execution analysis

Application

  • Desktop interfaces and charting
  • Local model services
  • Decision-flow visualisation
  • Operational logging and diagnostics

Project highlights

  • Decision visibility

    Inspect the path from signal to trade

  • Local forecasting

    Model services connected to the desktop

  • A documented evolution

    Analysis tools through to a richer research terminal

  • Research and paper trading

    Experiments kept distinct from a verified live record