imst-quant

imst-quant is a research pipeline for sentiment-driven trading that treats “whose sentiment” as seriously as “what sentiment.” Instead of averaging Reddit posts into a single signal, it learns which authors are actually worth listening to.

data lake

Reddit and market data flow through a medallion architecture:

• Raw - unprocessed ingestion from the Reddit API and market data feeds.
• Bronze - cleaned, deduplicated, schema-validated.
• Silver - joined and enriched (author history, ticker mentions, price context).
• Gold - feature tables ready for modeling.

influence scoring with GNNs

Rather than trusting every post equally, a graph neural network scores each author’s influence using historical prediction accuracy and their position in a trust network (who engages with, agrees with, or amplifies whom).

# author_graph: nodes = authors, edges = interaction/trust signals
influence_scores = gnn_model(author_graph, node_features=historical_accuracy)
weighted_sentiment = raw_sentiment * influence_scores[author_id]

forecasting and backtesting

Influence-weighted sentiment feeds a forecasting model, and every strategy is evaluated through a backtester that models realistic transaction costs - slippage and fees, not idealized fills - so backtest performance isn’t inflated by trading frictions that don’t exist in the real market.

status

This is a research pipeline: the data lake, GNN influence scoring, and cost-aware backtester are built; ongoing work is on forecasting model quality and strategy evaluation across market regimes.