How we build models you can check.
SparkTrade's models are built and tested by quantitative researchers, with input from experienced institutional managers. This is how we work.
of training data, from 1999 to today
U.S. stocks above $200 million in market value
technical, structural, fundamental and sentiment inputs
The data
Training data runs from 1999 to today, through several bull markets, bear markets and volatility spikes. Prices, fundamentals and technical features are rebuilt as they were known on each date, with controls for survivorship bias and look-ahead leakage.
Many inputs, not one factor
Each model weighs more than 300 technical, structural, fundamental and sentiment inputs. Instead of relying on a single factor, it looks for the strongest evidence of relative opportunity under current market conditions.
Short-term to multi-week horizons
Models cover horizons from short-term swing trades to multi-week positions, so each partner can choose a model that fits how its clients invest.
Regimes
Markets behave differently when they are calm and when they are stressed. Our models are trained and weighted by volatility regime, so patterns seen in quiet markets are not applied unchanged when volatility jumps.
Fixed factor directions
For each input, we fix whether a higher value may raise or lower the score. With other inputs held constant, that direction cannot reverse within the model. Other inputs can still change the overall score.
Testing on unseen periods
Models train on one window of history and are tested on the next. This process rolls forward through bull markets, bear markets and volatility spikes, with overlapping periods removed to prevent leakage. We judge models on risk-adjusted measures such as the Sharpe and Sortino ratios, after costs and slippage, rather than on hit rates alone.
Dated, versioned and labeled
Every score records the date it applies to and the model version that produced it. Sample, illustrative and research outputs are labeled as such, so they are not mistaken for production signals.
Where language models fit
Our signals come from statistical models, not chatbots. In Event Pressure, a language model drafts company research when prediction-market odds move or major news breaks, and a second pass checks each finding against its sources.
What we leave to you
SparkTrade does not place trades, hold client assets or tell anyone what to buy or sell. Partners and their clients make every decision.
Want the full methodology notes?
We share detailed model documentation with partners and data clients under NDA.
