python · data engineering

alternative data platform — indian equities

find information that moves stocks in the physical world, before it hits financial statements.

view source on github
python alternative data medallion architecture point-in-time feature store fastapi prefect timescaledb backtesting quant

what it is

most market signals show up in a company's numbers after the fact. alternative data is the bet that you can see the same thing earlier — in power demand, shipping, footfall, satellite imagery — if you can structure it. this platform is built to do that for indian equities.

it's a scalable skeleton with hard layer contracts plus one fully working vertical slice (posoco power demand). adding the other sources is a plug-in, not a rewrite — copy a template, implement three methods, add a feature recipe, and nothing downstream changes.

the medallion pipeline

→sources → ingest — pluggable data sources behind one abstract contract
→bronze — raw, untouched, in object storage (minio)
→silver — normalized records in postgres
→gold — a point-in-time feature store
→signals → backtest → delivery — factor models, an indian cost model, fastapi + grafana

the rule that runs through everything

point-in-time correctness is structural, not a checkbox. every feature carries (feature_date, published_date, as_of_date), and every read goes through one function that filters to what was actually knowable at the as-of moment. there's simply no api to see the future — which is the single most common way a backtest lies to you.

move the as-of date

a backtest on a given day may only use what was published by then. drag the date. rows that weren't known yet get struck through. that is lookahead bias, made visible.

featureaboutpublished
daily power demand2 july4 july
monthly auto salesjune10 july
cpi inflationjune14 july
quarterly results, company aapril to june18 july
quarterly results, company bapril to june25 july
daily power demand, revised2 july30 july

an illustrative month; the rows are made up to show the rule, not real releases.

pythonlanguage
posocoworking vertical slice
minio + pgstorage
fastapidelivery
i wanted to build the boring, correct part first — the part that stops you fooling yourself — and only then add sources. the anti-lookahead feature store is the whole point.

built by dharun ashokkumar