Methodology

Production pipeline principles

1. Point-in-time ingestion — version source records and preserve availability timestamps.
2. Schema and quality checks — reject impossible, duplicated or stale observations.
3. Leakage-safe features — fit transformations inside each training window.
4. Versioned inference — log model, feature, source and forecast timestamps.
5. Monitoring — track residual drift, calibration and failures by segment.

Production deployment should add source provenance, restatement handling, model cards, incident procedures and independent performance review.

Reported metrics and variables

Definitions for the figures shown across Performance, Reported Results, Forecast Signals, Company Research and ETF Coverage — every one is computed from real reported or forecast values, never estimated for display.

Forecast accuracy

1. MAE (mean absolute error) — average of |actualEPS − forecastEPS| across forecasts. The headline accuracy figure; lower is better, in dollars of EPS.
2. RMSE (root mean squared error) — sqrt(mean(error²)). Penalises large misses more heavily than MAE, so a RMSE well above MAE flags a few big outlier errors rather than uniformly noisy ones.
3. Median absolute error — the middle value of all absolute errors. More robust than MAE to a handful of extreme outliers (e.g. a one-off accounting item).
4. MASE (mean absolute scaled error) — MAE divided by the previous-quarter-naive baseline's own MAE over the same forecasts. Below 1.0 means the model beats simply assuming no change quarter over quarter; above 1.0 means the naive guess would have done better.
5. SMAPE (symmetric mean absolute percentage error) — mean(2·|error| / (|actual|+|forecast|)) × 100. A bounded percentage-error measure; can still swing widely for companies reporting EPS near zero, since any error is large relative to a tiny denominator.
6. Tolerance accuracy (within $0.01 / $0.05 / $0.10, within 5% / 10%) — the share of forecasts landing inside that absolute-dollar or percentage tolerance of the actual result.

Interval calibration

1. Prediction interval level — the target coverage the lower and upper bounds are built for, currently 60%: the actual is expected to land inside roughly three times in five, and outside the other two. It is a design target across many forecasts, not a promise about any single one. It was 80% until we measured what that cost — a median range of $0.805 against a median absolute error of $0.183, wide enough that over a third of companies carried no usable verdict. Narrowing it is a smaller claim honestly labelled, not a more accurate model; the forecasts and their errors are unchanged.
2. Interval coverage — the empirical share of actual results that fell within the published lower/upper bound, measured after the fact. Tracking close to the prediction interval level over many forecasts indicates well-calibrated (not over- or under-confident) intervals.
3. Average interval width — mean of upperBound − lowerBound. Narrower is more decision-useful, but only alongside healthy coverage — a narrow interval that misses often is worse than a wider one that's reliable.

Growth and direction

1. Expected QoQ growth — (forecastEPS − previousQuarterEPS) / |previousQuarterEPS| × 100, versus the company's own most recently reported quarter.
2. Expected YoY growth — the same calculation against the same fiscal quarter one year earlier, which controls for normal seasonal patterns within a company's business.
3. Directional EPS growth accuracy — the share of forecasts that correctly called the sign of EPS change versus the prior quarter (grew vs. shrank), independent of how close the exact figure was.

Confidence and signal labels

1. Expected change vs prior-year quarter — the forecast against the same fiscal quarter twelve months earlier. This is the like-for-like reading: it compares a quarter with its own counterpart, so a seasonal business is measured against its own season. Shown as "YoY" where space is tight, and it is the figure signal strength is built on.
2. Expected change vs previous quarter — the forecast against the quarter just reported. It moves with a company's seasonal cycle as much as its trajectory, so a large value often reflects the calendar rather than the business: a retailer's Christmas quarter against its autumn one can read as several hundred per cent without anything having changed. Useful for spotting sequential turns in businesses that are not seasonal; weaker than the prior-year comparison everywhere else.
3. Confidence score (0–100) — how narrow the prediction interval is relative to the point forecast, on a scale where a wider interval scores lower. Scored identically for an upcoming forecast and for every reported quarter, so a live forecast and the track record beneath it are directly comparable. Confidence label: High ≥ 56, Moderate 36–55, Low < 36. The cuts were re-fitted when the interval moved to 60% — a narrower interval raises every score, so leaving them would have promoted a third of companies a band without the model being any better. They now hold the same share of companies, with the same realised error, as before the change. The bands look low against a 0–100 scale because most forecasts carry a genuinely wide interval; the performance page publishes how far each band actually missed.
4. Signal strength — a plain-magnitude label on expected YoY growth: Strong at 20% or more (either direction), Moderate from 8–20%, Weak below 8%. It describes the size of the expected move, not the model's confidence in it — a Strong signal can still carry Low confidence.

Baselines (Model Performance page)

1. Previous quarter — naive persistence: assumes next quarter's EPS equals the quarter just reported.
2. Same quarter, one year earlier — seasonal-naive: assumes next quarter matches the same fiscal quarter a year ago.
3. Four-quarter moving average — the average of the last four reported quarters' diluted EPS.
4. Sector-quarter historical median — the median diluted EPS across every same-sector, same-fiscal-quarter observation available up to that point in time.
5. Model baseline win rate — the share of forecasts where the model's absolute error was smaller than the best-performing naive baseline available for that same observation.