Signal Methodology
Model: v5.7 + v5.8.1 ensemble · Last updated: June 2026
WatchSignal generates AI-powered trading signals entirely on your device. This page explains how those signals work, what data they use, and what their limitations are. We believe transparency builds trust — so we've published this openly. For the headline accuracy numbers and how we validate them, see our accuracy page.
How Signals Work
Each signal is a three-class prediction: Bullish, Neutral, or Bearish.
| Signal | Meaning |
|---|---|
| Bullish | The model sees bullish conditions in the stock's recent history |
| Bearish | The model sees bearish conditions |
| Neutral | The model has no confident directional view — staying flat is the read |
Every signal comes with two additional values:
- Confidence (0–100%) — how concentrated the model's probability is across the three classes. Lower confidence means the signal is closer to neutral.
- Signal score (−1 to +1) — a continuous measure of directional conviction. Scores near 0 are effectively neutral regardless of the directional label.
A Bullish signal with 55% confidence is very different from one at 88%. Always factor confidence into your interpretation.
What the Model Analyses
The model builds a feature vector per stock from price action, market context, earnings history, and the macro environment. Different heads consume different slices of that vector — the directional heads use 70 features, the macro-regime head uses 121, and the neural head uses a 41-feature core. They share the same underlying signals, described below.
Price & Technicals
- Open, High, Low, Close, and Volume (normalised ratios — not raw prices)
- RSI (14-day), MACD and signal line, Bollinger Bands %B
- Average True Range, volume ratio, On-Balance Volume trend
- Stochastic %K, SMA50/SMA200 ratio, overnight gap, intraday range, true-range ratio
Weekly Context
- Most recent weekly OHLCV candle aligned to each daily bar — gives the model a broader view of trend
Market Regime
- SPY 20-day return and relative ATR — tells the model whether the broader market is trending or volatile
- VIX (CBOE Volatility Index) — captures the implied-volatility environment. The VIX level also selects which ensemble blend weights are used (see The Model)
Macro Stress
- Yield-curve slope (10Y minus 2Y), investment-grade credit spread (BAA−AAA), and corporate option-adjusted spread — each z-scored. These tell the model how stressed the broader financial system is, independent of any single stock.
Relative Strength
- 5-day and 20-day ticker return relative to SPY — how this stock is performing versus the market
Earnings
- Proximity to the next earnings date
- Post-earnings 1-day and 5-day returns across recent quarters
- EPS surprise magnitude and a multi-quarter surprise trend (180-day decay)
- Pre-earnings drift average and standard deviation
These features help the model recognise typical earnings run-up and post-announcement patterns specific to each stock.
Macro Events
- Proximity to scheduled FOMC, CPI, and NFP release dates
- The stock's historical sensitivity (beta and R²) to those macro events, computed over a trailing window
Dividends
- Proximity to the next ex-dividend date
Price Momentum
- 3-month, 6-month, and 12-month momentum (z-score normalised)
- 12-month momentum uses the skip-last-month convention to avoid short-term reversal bias (Jegadeesh & Titman, 1993)
Price Position
- Distance from 52-week high and 52-week low (z-score normalised, George & Hwang, 2004)
Market Breadth
- Percentage of S&P 500 stocks above their 200-day and 50-day moving averages
The Model
WatchSignal does not use a single model — it uses an ensemble specialised by prediction horizon. For each horizon (short ~5 days, medium ~10 days, long ~21 days, and a macro/regime head at ~63 days) two complementary models vote:
- Gradient-boosted decision trees (XGBoost). Four classification heads — one per horizon — read the full feature vector. A parallel set of regression heads produces the numeric price targets shown on the Forecast tab.
- A compact neural network with per-ticker embeddings. A small multi-layer network (41 inputs + a 16-dimensional learned embedding per ticker → 256 → 128 → 64 → three classes) lets the model learn name-specific behaviour that the trees alone miss. Unseen tickers fall back to an averaged "cold-start" embedding.
Regime-conditional blending. The two models' calibrated probabilities are combined per horizon, with blend weights that depend on the current VIX regime (low / mid / high volatility). The weights are validated and shipped in the model bundle rather than hand-tuned. The model was trained on a universe of roughly 390 US-listed equities and ETFs.
On-Device Inference
The model runs entirely on your phone using ONNX Runtime Mobile. No price data, portfolio holdings, or analysis results are ever sent to WatchSignal's servers.
- iOS: Uses Apple's CoreML execution provider, which routes computation to the Neural Engine or GPU where available, with a CPU fallback.
- Android: CPU inference.
- Typical inference time: under one second.
Each model head is downloaded once and verified against a SHA-256 checksum before use. Heads are never modified remotely without your app updating.
Calibration
Raw model outputs are calibrated per head using isotonic regression, so that a stated confidence corresponds to the empirical hit rate — a 70% confidence score should mean the model is correct roughly 70% of the time, not more, not less. The numeric price ranges on the Forecast tab use conformal prediction (a 90% coverage target), which gives an honest interval rather than a single point guess. We monitor calibration in production and ship a recalibration if it drifts.
Personalisation (Optional)
Signal+ subscribers can correct signals they disagree with. When you do, a small on-device calibration layer adjusts slightly toward your feedback without retraining the full model. This personalisation is stored locally on your device and is never shared.
Data Sources
| Data | Source | Delay |
|---|---|---|
| Daily OHLCV candles | Yahoo Finance | Up to 15 minutes delayed |
| Live quote (price, % change) | Yahoo Finance | Real-time during market hours |
| Analyst ratings | Yahoo Finance | Trailing 12 months rolling |
| Earnings dates & surprises | Yahoo Finance | Near real-time (dates); delayed (results) |
| VIX, SPY | Yahoo Finance | Real-time during market hours |
| Dividend dates | Yahoo Finance | Near real-time |
| Macro stress (yield curve, credit spreads) | FRED (Federal Reserve) | Daily |
| Trending movers & news | Polygon.io (Yahoo Finance fallback) | Real-time; cached 5 minutes |
| Macro event calendar (FOMC, CPI, NFP) | Bundled in-app calendar | Compiled at release; updated each app version |
Data may be delayed, incomplete, or temporarily unavailable. See our Terms of Service for full data liability terms.
Accuracy & Limitations
Accuracy varies by horizon. On out-of-sample walk-forward validation, the published per-horizon scores range from about 45% to 59% on the three-class (Bullish / Neutral / Bearish) problem — where a coin flip is roughly 33%, not 50%. The exact per-horizon numbers, and the validation method behind them, are on the accuracy page. In plain terms:
- The model is frequently wrong — directional accuracy in the high-50s still means a large minority of calls miss.
- Past backtested accuracy does not guarantee future performance.
- Accuracy varies significantly by stock, sector, and market regime.
- ETFs receive reduced signal quality because earnings, analyst, and dividend features default to neutral.
- Stocks with limited available history may produce lower-quality signals.
- The model does not have access to material non-public information, breaking news, or corporate actions not yet reflected in price.
- Aggregate accuracy is not per-ticker accuracy. The in-app Forecast Accuracy tab on each ticker shows that name's actual hit rate.
Signals are a starting point for research — not a conclusion.
Not Financial Advice
WatchSignal signals are informational outputs only. They are not personal investment recommendations, and WatchSignal is not a registered investment adviser under SEC, FCA, MiFID II, or any other regulatory framework.
Before making any investment decision, you should:
- Consider your own financial situation, risk tolerance, and investment goals
- Consult a qualified financial adviser if appropriate
- Read the full risk disclosures in our Terms of Service
Investing in securities involves risk. You may lose some or all of your investment.
Questions?
If you have questions about our methodology or want to report an issue, contact us at support@watchsignal.ai.