Upside
Finance / Market AnalyticsWeb & PWA

Upside

Market data and forecasting platform that turns prices into calibrated probabilities across five markets, with an AI analyst whose every call is recorded and scored.

Project Overview

Upside collects prices from several market data providers, rebuilds them into clean daily sessions, and estimates the probability of a rise, a flat move or a fall over the next 1, 5, 20 and 60 sessions — for spot gold, the major currency pairs with a self-computed dollar index, large-cap crypto, Vietnam's VN30 and a watchlist of US stocks. Forecasts come from three layers that compete in the open: a base rate, a set of technical rules, and LightGBM models validated walk-forward with purging and embargo, so no feature ever sees data from after the session it describes. Probabilities become signals with a stop loss, a take profit and an expiry, and a paper-trading simulator follows every signal after spread and fees. The AI layer runs on Claude through the Agent SDK. It scores incoming news headlines into symbol tags and impact direction, writes plain-language explanations of each signal and a morning briefing sent to Telegram, and acts as an AI analyst: every session it reads a point-in-time snapshot of prices, indicators, the other strategies' views and the news, and records its own probabilities, which are scored exactly like every other strategy. Its record is live only — never backfilled — because a language model has already seen the history a backtest would replay. Every number in AI-written text is checked against the data it was given; text that cites a figure not in the input is rewritten once, then dropped. The whole system is built to be checked: every forecast is stored before its outcome is known and can never be edited, live results are never mixed with backtests, and a strategy only goes live after beating the layer below it out of sample, after costs. In walk-forward tests on gold no technical rule consistently beat the base rate, so the base rate remains the strategy Upside trusts — and the public site says so. Upside is our own product and currently an internal tool, with the public site in preview for expert review; we designed, built and operate it end to end.

Our Contribution

  • Product design and full-stack development (in-house product)
  • Market data pipeline across several providers, with cross-source quality checks
  • Forecasting engine: base rate, technical rules and walk-forward LightGBM for four horizons
  • Point-in-time feature store with tests that reject any look-ahead
  • AI analyst on the Claude Agent SDK, recording its own scored probabilities every session
  • AI news scoring, signal explanations and morning briefings, checked number by number
  • Append-only forecast ledger, automatic scoring and a promotion gate between strategy layers
  • Paper-trading simulator with spread and fees, and an equity curve per strategy
  • Market-aware scheduler for New York, UTC and Hanoi sessions, with monthly retraining
  • Web app and PWA with Telegram and web push alerts; bilingual public site
  • Production infrastructure, nightly backups with restore drills, and ongoing operations

Tech Stack

PythonpandasLightGBMAPSchedulerClaude Agent SDKNestJSPrismaPostgreSQLRedisReactTypeScriptAnt DesignLightweight ChartsZodWeb Push

Key Features & Highlights

Probabilities of a rise, a flat move or a fall for 1, 5, 20 and 60 sessions — not single-point predictions

Five markets live: gold, major currencies with a computed dollar index, crypto, VN30 and US stocks

AI analyst records its own probabilities every session and is scored like any other strategy — live only, never backfilled

Claude scores news headlines into symbol tags and impact direction

AI-written explanations and briefings checked number by number against the data behind them

Every forecast stored before its outcome is known and never edited; live results never mixed with backtests

Walk-forward validation with purging and embargo, and tests that reject look-ahead features

A strategy goes live only after beating the layer below it out of sample, after costs

Paper-trading simulator follows every signal after spread and fees

Price charts with a forecast cone for each horizon

Trading ideas from the system, the AI analyst and invited experts, all scored automatically

Telegram and web push alerts, and a bilingual English and Vietnamese site

App Screenshots

Interested in working with us?

We'd love to help you
build your next product.

Whether you have a clear brief or just an early idea — reach out below and our team will get back to you within one business day.