Speculative ideas and research directions not yet assigned to a project. Items here are possibilities, not commitments.
Store Ansible playbooks in the notes system so an agent can read, update, and redeploy them centrally without needing filesystem access to the repo. Two approaches worth exploring:
• Notes as source of truth: store playbooks as codeblocks in notes; agent edits notes directly and a sync script pushes to the filesystem before running. Clean but diverges from git history.
• Git-backed: agent clones the ansible repo, edits files, commits and pushes. Notes hold documentation/reference only. Keeps full git history and works with existing tooling. Probably the better approach — could trigger via a runnable sheet.
→ Feasibility assessment with prior work, data-granularity analysis and references: ideas/spectrogram-trading.
Concept: Apply spectrogram (short-time Fourier transform or wavelet transform) to price/volume time-series to expose periodic structure in the frequency-time plane, then feed the resulting 2-D image to a pattern-recognition model.
Motivation: Market data has multi-scale cyclic behaviour (intraday rhythms, weekly patterns, regime changes) that is largely invisible to scalar technical indicators. A spectrogram converts these into spatial features that convolutional networks are well-suited to detect.
• Compute STFT (or CWT) over a sliding window of a price or return series. Output: 2-D array (frequency × time).
• Treat spectrogram tiles as images — normalise, optionally log-scale the magnitude.
• Feed to a CNN (or ANN with even convolutions) trained to predict direction, regime, or risk level.
• Even (symmetric) convolution kernels preserve phase information and avoid directional bias — worth exploring as an inductive bias for this domain.
• Python: scipy.signal.stft or pywt (CWT). Visualisation via matplotlib.pyplot.specgram.
• Label generation: forward returns, volatility regimes, or labelled trend/range periods.
• Consider multi-channel input: price, volume, spread each as a separate spectrogram channel.
• Risk of lookahead bias: ensure STFT window only uses past data relative to the label timestamp.
• Wavelet coherence for pairs/correlation trading.
• Audio ML literature (spectrogram CNNs for speech/music) is directly applicable — large body of prior art to draw on.
• Could integrate with Envoy for automated periodic analysis delivery.