Trend analysis

Take any numeric time series — a stock's daily close, a FRED macro indicator, a treasury yield history — and run it through the full quantitative workup: descriptives, moving averages, trend line, outliers, optional correlation against a benchmark, and a deterministic forecast forward with a 95% prediction interval. Everything an analyst writes a notebook for, in one chain of cheap calls.

When to use this pack

You have a question like "is AAPL trending up over the last year — and what does the next quarter look like?" or "is unemployment a leading indicator for fed-funds moves?" and want a deterministic numerical answer (slope, r², outlier dates, point forecast + 95% interval) instead of a hand-wavy LLM summary or hallucinated projection. The stats + forecast steps are pure-CPU and free over PoW; only the upstream data fetch (finance/macro) is paid.

Tools in this pack

• Stock historical bars $0.015 GET /api/stock-history Historical OHLCV bars for a symbol. Configurable interval (1m, 5m, 15m, 30m, 60m, 1d, 1wk, 1mo, 3mo) and range (1d, 5d, 1mo, 3mo, 6mo, 1y, 2y, 5y, 10y, ytd, max). Intraday intervals are limited by Yahoo to ~60 days of data. Returns a flat array of bars (time, open, high, low, close, volume) ready for charting or backtests.
• FRED time series $0.015 GET /api/fred-series Fetch any of FRED's ~800,000 economic time series by series ID — GDP (GDPC1), CPI (CPIAUCSL), unemployment (UNRATE), fed funds (DFF), and so on. Supports date windowing and the standard FRED units transformations (lin, chg, ch1, pch, pc1, pca, cca, log). ?seriesId=GDPC1&startDate=2018-01-01&endDate=2023-12-31&units=pc1
• Stats summary $0.001 POST /api/stats-summary Compute the full descriptive-stats panel for an array of numbers in one call: count, sum, mean, median, mode, stddev (sample), variance, min, max, range, q1, q3, IQR. Beats calling 12 separate tools when you already have the array in front of you.
• Moving average (SMA + EMA) $0.001 POST /api/moving-average Compute simple (SMA) and exponential (EMA) moving averages over a numeric series. Returns one value per input position — the first (window-1) SMA values are null since there isn't enough history. EMA uses the standard alpha = 2/(window+1) smoothing factor used in technical analysis.
• Linear regression (OLS) $0.001 POST /api/linear-regression Fit a least-squares line y = slope·x + intercept to two equal-length series. Returns slope, intercept, r² (variance explained), and optionally predicted y values for new x inputs — useful for trend extrapolation (e.g. project next quarter's revenue from the last 8 quarters).
• Outlier detection (IQR + z-score) $0.001 POST /api/outliers Flag outliers in a numeric series using either the IQR rule (Tukey fences at 1.5·IQR — robust, default) or z-score (|z| > threshold — assumes normality). Returns the outlier values + their indices + the thresholds used so you can decide whether to trust them.
• Correlation (Pearson) $0.001 POST /api/correlation Pearson correlation coefficient between two equal-length numeric series. Returns r (the correlation, -1 to 1), r² (variance explained), n (sample size). Use this to ask things like: is a stock's daily return correlated with a macro indicator? Are two FRED series moving together?
• Forecast backtest (MAPE + RMSE) $0.001 POST /api/forecast-eval Backtest a forecasting method on the input series by holding out the last `testSize` observations, forecasting them, and computing MAPE (mean absolute percentage error) + RMSE (root mean squared error). Lets an agent pick which method (mean / naive / drift / ses / holt / holt-winters) actually fits its data before committing to a forward forecast. Always returns a `warnings` array — empty when the backtest is well-posed, populated when `testSize` exceeds n/2 (treat error as indicative not predictive).

Workflow

1. Fetch the series. For an equity ticker, call stock-history with range=horizon (or "1y" if unspecified) and pull the array of `close` prices in chronological order. For a macro indicator, call fred-series with the series id (UNRATE, CPIAUCSL, FEDFUNDS, etc.) and pull the array of `value`s.
2. Run stats-summary on the values to get the full descriptive panel (mean, median, stddev, min, max, q1/q3, IQR). This is the one-line "what does this series even look like" answer — agents that skip this step end up reporting trends without context.
3. Smooth the noise with moving-average. A 20-day SMA is the textbook short-term trend smoother for daily prices; a 12-month MA suits monthly macro data. Use which="both" so you can compare SMA (lagging but stable) with EMA (responsive but jittery).
4. Fit linear-regression with x = [0, 1, ..., n-1] (just the index) and y = values. Slope tells you direction + magnitude per unit time; r² tells you how clean the trend is (>0.7 = strong trend, <0.3 = mostly noise). Pass `predict` for next-N-period extrapolation if the user wants a projection.
5. Flag anomalies with outliers method="iqr" — Tukey fences (1.5·IQR) are the conservative default. Report the indices + values; agents should then map indices back to dates from the original fetch so the answer says "2024-03-14: $187.23 outlier" not just "index 142".
6. If the user asked a comparison question ("is AAPL correlated with the S&P?", "do CPI and fed funds move together?"), repeat steps 1-2 for the benchmark series, then call correlation with the two equal-length arrays. r above 0.7 = strong same-direction move; near 0 = independent; negative = inverse. Use the `interpretation` field as your one-line answer.
7. Pick a forecast method honestly by backtesting. Call forecast-eval three times — once each with method="drift", "ses", "holt" — passing the same values + testSize (≈ 20% of the series, capped at half). Compare RMSE; the lowest wins. Check `warnings` — non-empty means treat the result as indicative not predictive. Skip the bake-off only if you already know the series shape (e.g. holt-winters for clearly seasonal data with a known period).
8. Forecast forward with the winning method. Call forecast-naive / forecast-ses / forecast-holt (whichever won) with the full values + the user's horizon. Return the point forecast AND lower95/upper95 — never report a point estimate without its interval; that's the whole reason these tools exist instead of an LLM guess. Combine summary + trend + outliers + optional correlation + forecast into a single JSON object. That's the deterministic analyst-grade reply.

Run it in Claude

claude mcp add agent402 -s user -- npx -y agent402-mcp@latest

Then paste this prompt into Claude:

Run a full trend analysis on AAPL over the last 1y using Agent402, then project the next quarter forward. (1) Fetch the daily closes via stock-history (ticker=AAPL, range=1y). (2) Run stats-summary on the closes for the descriptive panel. (3) Run moving-average with window=20, which="both" — compare SMA vs EMA. (4) Run linear-regression with x=[0..n-1], y=closes; report slope (annualized = slope·252), intercept, r². (5) Run outliers method="iqr" and map the flagged indices back to actual dates from the fetch. (6) Pick a forecast method: call forecast-eval three times with method="drift", "ses", "holt" and testSize=50 (≈ 20% of a 252-day year); pick the lowest RMSE. (7) Forecast the next ~63 trading days using the winning method (forecast-naive / forecast-ses / forecast-holt) and report both point and 95% interval. (8) Return a single JSON object: {summary, trend, outlierDates, forecastMethod, forecastWithIntervals, oneLineConclusion}. The stats + forecast steps are free over PoW; only the stock-history fetch is paid.

← All skill packs