Published without the client's name. We name a client only with their agreement.

Market Prediction Platform

Turning one analyst's chart-reading method into a machine-learning platform

Our client's work rests on a detailed personal system for reading the markets, built up over years of practice. CodeLeap turned that method into supervised machine-learning models trained on the analyst's own labelled history, then built the platform around them — a daily pipeline, per-indicator accuracy validation and an API through which the predictions can be served and built upon.

Market Prediction Platform — Machine learning case study

One analyst's eye, scaled.

Our client is a financial analysis firm whose work rests on a detailed personal system for reading the markets, built up over years of practice. Every day an analyst would work through the same charts, reading trend patterns, moving averages and a set of oscillators, and marking each one with a direction by eye. The method was proven and the judgement behind it was valuable, but performing it manually was slow, repetitive and impossible to scale beyond what one person could watch. CodeLeap partnered with the firm to turn that method into supervised machine-learning models trained on the analyst's own labelled history, and then to build the platform around them, comprising a daily time-series pipeline, per-indicator accuracy validation and an API through which the predictions can be served and built upon.

Product features

Trend and Oscillator Models

Supervised models trained on the analyst's own labelled history, learning to read price action, moving averages and each individual oscillator the way the analyst would read them by eye. Training on the client's own labels rather than on a generic definition of each signal was central to the approach, because the commercial value of the system lies in reproducing their particular interpretation rather than a textbook one.
Daily Time-Series Pipeline

A pipeline that pulls in each day's price and indicator data, runs it through the models and writes the resulting labels straight back into the client's working spreadsheet, so the daily read-and-label routine now completes itself before anyone sits down to it. Removing the manual copying between systems also removed a whole category of transcription error from a process where a single mislabelled indicator changes the conclusion.
Per-Indicator Accuracy Validation

The models' labels are measured against the analyst's own, indicator by indicator rather than as a single aggregate score, so accuracy is demonstrated before any output is relied upon. This granularity matters in practice, since it shows precisely which readings the models have mastered and which still warrant a human eye, and it gives the client a defensible basis for trusting the system.
Prediction API

Predictions are served through a clean, well-documented API, allowing the client to build their own products and workflows on top of the platform without touching the models underneath. This turns what began as an internal efficiency exercise into a capability the firm can commercialise, with the separation between the API and the model layer meaning the models can be retrained and improved without disrupting anything built above them.
Time-Series Data Processing

A specialised preprocessing layer handles the volume, noise and non-stationarity inherent in market data, cleaning and shaping each day’s inputs so the models receive reliable features rather than raw feeds. Much of the accuracy of the finished system comes from the care taken at this stage, before any model sees the data.
Compute Cost Optimisation

The infrastructure was architected throughout to balance predictive accuracy against computational cost, so that the platform delivers strong results without the cloud spend that continuous market data processing can easily generate. For a system intended to run every day indefinitely, keeping the economics sound was as much a requirement as the accuracy itself.

Challenge

The client had a method that worked and an analyst who applied it consistently, but the method could only ever move at human speed. Reading trend patterns, moving averages and a full set of oscillators off the charts each day, marking each with a direction and recording the results, consumed hours of skilled attention on work that was fundamentally repetitive, and it capped the number of instruments the firm could realistically cover. Automating it was not a matter of applying standard technical analysis, because the value lay in this analyst's specific interpretation of each signal, which meant the models had to learn from their labels rather than from a general rule. The underlying data compounded the difficulty, since financial time series are noisy, non-stationary and shaped by countless external factors, making this one of the more demanding applications of machine learning. The firm needed a platform that could reproduce their judgement faithfully enough to be trusted, prove that it was doing so, run reliably every day, and remain affordable to operate as coverage grew.

Solution

CodeLeap converted the client's chart-reading method into a set of supervised machine-learning models trained directly on the analyst's own labelled history, then built the platform that puts them to work. Each day the pipeline gathers the latest price and indicator data, runs it through the models and writes the labels back into the client's working spreadsheet, so the routine that once occupied an analyst's morning now completes on its own. Accuracy is validated indicator by indicator against the analyst's own readings, which means confidence in the system is evidenced rather than assumed, and the same models are exposed through an API that the client can build products on. The architecture balances predictive performance against compute cost so that the platform remains commercially viable as data volumes and coverage increase. The outcome is expert judgement scaled well past what a single analyst could ever watch, applied consistently every day, and available as a service the firm can develop further.

Tech stack

The tech chosen was our speciality, enabling fast development of performant machine learning models and a reliable infrastructure.

Backend: Python, Django, PostgreSQL
Machine Learning: Custom-trained supervised models, technical indicator analysis, per-indicator accuracy validation
Hosting: AWS