How Upside builds data and forecasts

1. Collecting prices

Upside collects hourly candles for spot gold (XAU/USD) from Twelve Data and a daily price snapshot from Alpha Vantage as a second source. Older history comes from Twelve Data daily candles, whose closing time is approximate and is marked as such.

2. Building daily sessions

A trading day runs from 17:00 New York to 17:00 New York the next day, including daylight saving changes. Upside builds each daily candle from the hourly candles in that window and drops hours when the market is closed.

3. Checking quality

• Invalid candles (for example a high below the close) are marked bad and never used.
• A jump larger than six times the recent volatility that the second source does not confirm is marked suspect.
• Differences between sources above a threshold are logged for review, without overwriting either source.
• When a provider revises an old candle, the previous values are kept in a revision log rather than overwritten.

4. Defining outcomes

For each horizon (1, 5, 20 and 60 sessions) the outcome is a rise, a flat move or a fall. A move counts as flat when it stays inside a dead zone that scales with volatility: a quarter of the daily average true range, multiplied by the square root of the horizon.

5. Producing probabilities

Strategies are arranged in layers. Layer 0 holds baselines: the base rate (how often each outcome happened in the past ten years) and momentum. Layer 1 holds technical rules such as moving-average trend, RSI, MACD, ADX, Bollinger Bands and Donchian channels, plus an ensemble of them.

Each rule’s raw score is turned into probabilities by a calibrator fitted walk-forward on a rolling ten-year window, with a gap between training data and the forecast date so that no future information leaks in.

6. Choosing which strategy to trust

A strategy is trusted (the champion) only if its Brier skill beats the base rate out of sample on both the full history and the period since 2020. Today no technical rule passes that test on gold, so the base rate is the champion. Changing the champion is a recorded decision, not an automatic switch.

7. Recording and scoring

Forecasts are stored before the outcome is known, and the database itself rejects any later change or deletion. When the horizon ends, each forecast is scored. Live results are never mixed with backtests, and forecasts filled in after the fact are marked and excluded from the live record.

8. Limits

• Probabilities describe the past behaviour of prices; markets change and past skill may not persist.
• A backtest, even walk-forward, is easier than live forecasting. Trust the live record more.
• Upside does not use news, positioning or order flow in its current models.