Real-Time Time-Series Signal Analysis Platform

A real-time signal visualization and quantitative analysis platform - originally built for financial time-series, the architecture and signal processing primitives apply directly to any streaming time-series data including physiological and neural signals.

The dashboard ingests a live stream of samples, applies a battery of DSP-style filters and statistical estimators, and renders the results - including a Welch power-spectral-density panel and an STFT spectrogram - in a single Streamlit interface. Every indicator is implemented from scratch on top of NumPy / SciPy / Pandas so its filter structure is visible, not hidden inside a black-box library.

(Originally prototyped as a stock-price toy for my grandmother, who loves investing.)

---

Signal Processing Framework

Every "technical indicator" in this dashboard is a digital signal processing primitive applied to a one-dimensional time-series. The financial vocabulary is a thin wrapper around standard DSP operators:

Indicator	DSP interpretation	Formulation
SMA (Simple Moving Average)	Finite impulse response (FIR) box filter - convolution with a uniform 1/N kernel of length N. Linear phase, sinc-shaped magnitude response.	y[n] = (1/N) · Σ x[n-k] for k = 0..N-1
EMA (Exponential Moving Average)	First-order infinite impulse response (IIR) filter with an exponential decay kernel. Single real pole; geometric weighting of past samples.	y[n] = α·x[n] + (1-α)·y[n-1], α = 2/(N+1)
RSI (Relative Strength Index)	Bounded momentum oscillator - half-wave rectification of the first difference, followed by Wilder IIR smoothing of the positive and negative branches.	RS = avg_gain / avg_loss, RSI = 100 − 100/(1+RS)
Bollinger Bands	Volatility envelope - rolling mean ± k·σ over a sliding window. Statistical tolerance band on the recent local distribution.	(μ_t − k·σ_t, μ_t + k·σ_t)
MACD	Bandpass filter - difference of two EMAs at different time constants isolates a passband between them; the signal line is an EMA of that bandpass output.	EMA_fast(x) − EMA_slow(x), signal = EMA_sig(MACD)
ATR (Average True Range)	Rolling amplitude envelope - Wilder-smoothed true range; tracks instantaneous signal volatility.	`EMA(max(high-low,
OBV (On-Balance Volume)	Cumulative sign-modulated integrator - discrete analogue of a coulomb counter applied to volume signed by Δprice.	OBV_t = OBV_{t-1} + sign(Δp_t)·v_t
Rolling z-score	Local standardisation - mean-centred and scaled by rolling σ so heterogeneous channels can be compared on a common axis.	z_t = (x_t − μ_t)/σ_t
Welch PSD	Power spectral density via Bartlett-windowed segment averaging. The standard estimator for stationary spectral content.	`Ŝ(f) = (1/K) Σ
STFT spectrogram	Short-time Fourier transform - sliding windowed FFT producing a time × frequency energy surface.

Treating indicators this way makes their behaviour predictable: SMA introduces a known group delay of (N−1)/2, MACD has a passband determined by the two EMA time constants, Bollinger Band breaches are formal outliers under a local Gaussian assumption, etc. It also makes the codebase straightforwardly portable to any other streaming signal.

---

Extending to Other Domains

Although the dashboard ships with a Yahoo-Finance loader, none of the analysis modules know anything about prices. The exact same pipeline works for:

• EEG / LFP recordings. Replace fetch_stock_data(...) with an MNE / NWB reader. The Welch PSD becomes a band-power estimator (alpha 8-13 Hz, beta 13-30 Hz, gamma 30+ Hz). The STFT spectrogram is the canonical time-frequency view used in cortical analysis. Bollinger Bands become threshold envelopes for event detection on band-pass-filtered traces.
• Neural spike rates. Bin spike times into a rate signal, feed it in as Close. SMA produces the trial-averaged firing-rate envelope; EMA gives an online causal estimate. Autocorrelation surfaces refractoriness and rhythmicity. The HMM regime decoder is structurally identical to the behavioural-state decoders used in the honors-thesis neural-representation-explorer - same generative model, different domain.
• Physiological signals (HRV, EMG, respiration). RSI's rectify-and-smooth structure is the standard EMG-envelope pipeline. ATR captures rolling amplitude for burst detection. The rolling z-score is the normalisation step before any cross-channel comparison.

What changes when you swap domains: the loader, the units on the axes, and the interpretation of dominant frequency peaks.

What stays the same: every module under indicators.py, spectral.py, stats.py, and ui.py. The DSP is domain-agnostic.

---

Directory Structure

real-time-stock-dashboard/
├── stock_dashboard.py     # Streamlit entry point + data ingestion
├── config.py              # All hyperparameters in one place
├── indicators.py          # SMA, EMA, RSI, Bollinger, MACD, ATR, OBV, z-score
├── spectral.py            # Welch PSD, STFT spectrogram, band-power
├── stats.py               # ACF, returns distribution + KS, Sharpe, drawdown, HMM
├── ui.py                  # Plotly / Streamlit rendering helpers
├── requirements.txt
├── README.md
├── LICENSE
└── Example.png

Features

• Streaming ingestion with Yahoo-Finance as the default source; trivially swappable for any other sampled signal.
• Indicator overlays on candlestick or line charts: SMA, EMA, Bollinger volatility envelope.
• Dedicated panels for MACD (bandpass), ATR, rolling z-score, and OBV.
• Spectral analysis panel - Welch PSD with dominant-peak annotation + STFT spectrogram (the explicit bridge to EEG/neural pipelines).
• Statistical panels - rolling autocorrelation, returns distribution with normal-fit overlay and KS statistic, rolling annualised Sharpe, drawdown chart.
• Optional HMM regime decoder - 3-state Gaussian HMM on log-returns, decoded states drawn as background bands on the main chart.
• Live sidebar quotes for a configurable watchlist.

Installation

git clone https://github.com/peterajhgraham/real-time-stock-dashboard.git
cd real-time-stock-dashboard
pip3 install -r requirements.txt
python3 -m streamlit run stock_dashboard.py

Python 3.9+ is recommended. hmmlearn is optional - the HMM regime panel degrades gracefully if it is not installed.

Usage

Sidebar controls:

• Ticker - any symbol supported by Yahoo Finance.
• Window - analysis window (1d … max); the sampling interval is selected automatically.
• Chart type - candlestick or line.
• Overlay indicators - SMA, EMA, Bollinger Bands superimposed on the price chart.
• Panels - toggle MACD, ATR + z-score, OBV, Welch PSD, STFT spectrogram, autocorrelation, returns distribution, rolling Sharpe, drawdown, HMM regimes.

All hyperparameters (window lengths, smoothing constants, FFT segment sizes, HMM seed, colour palette) live in config.py.

Known Issues

• Data fetching errors: invalid tickers or rate-limited requests return empty data. The dashboard surfaces a Streamlit error.
• Short windows: spectral estimators and the HMM need a reasonable number of samples; they degrade to informational messages rather than crashing.

Contributing

Contributions welcome - particularly loaders for other signal domains (EEG, NWB, OpenBCI, EDF) so the platform can be exercised on physiological data without modification.

License

MIT - see the LICENSE file.

Contact

peter_graham@brown.edu